Breeding value prediction method and system based on multi-scale gene feature tensor

Abstract

This invention discloses a method and system for predicting breeding values ​​based on multi-scale gene feature tensors. The method obtains a preprocessed genotype matrix by performing quality control, missing genotype imputation, and standardization on whole-genome SNP marker data of the target population; it then uses a linear mixture model to correct for fixed effects on phenotypic observations; based on gene function annotation information and linkage disequilibrium structure information, SNP sites are divided into multiple functional regions, and a multi-scale genome feature tensor is constructed after dimensionality reduction using principal component analysis; a fused deep neural network model is constructed and trained, and the breeding value prediction results and prediction uncertainty quantification indicators are output through Monte Carlo Dropout inference. This invention can effectively capture local linkage disequilibrium patterns and long-range epistatic interactions, is applicable to endangered small population scenarios, and provides individual-by-individual prediction reliability assessments, providing a scientific basis for decision-making in the conservation and breeding of endangered local livestock and poultry genetic resources.

Description

Technical Field

[0001] This invention belongs to the field of gene breeding technology, specifically relating to a method and system for predicting breeding values ​​based on multi-scale gene feature tensors. Background Technology

[0002] Genomic selection (GS) is a core technology in modern plant and animal breeding. Its basic principle is to utilize high-density single nucleotide polymorphism (SNP) marker information covering the entire genome, and simultaneously estimate the effects of all marker sites using statistical or machine learning models. This predicts the individual's genomic estimated breeding value (GEBV), enabling early assessment and selection of the genetic potential of candidate individuals. Since Meuwissen et al. proposed the concept of genomic selection in 2001, this technology has achieved significant results in the breeding practices of commercial livestock and poultry breeds such as dairy cows, pigs, and chickens, greatly shortening generation intervals and accelerating genetic progress.

[0003] However, the practical application of whole-genome selection technology currently faces a significant and urgent bottleneck: the preservation and breeding of endangered local livestock and poultry genetic resources. Implementing whole-genome selection breeding for these endangered local breeds presents multiple technical challenges: the reference population is extremely small, and the predictive accuracy of traditional whole-genome selection methods (such as GBLUP, BayesA, BayesB, BayesCπ, etc.) is highly dependent on the size of the reference population; the importance of epistatic effects for local breeds is also a concern. During long-term natural selection and adaptive evolution, local breeds have accumulated a large number of epistatic interactions related to environmental adaptability, disease resistance, and quality traits. These nonlinear interactions between genes are a crucial genetic basis distinguishing local breeds from commercial breeds. However, traditional linear models (GBLUP and Bayesian methods) are inherently unable to effectively capture these non-additive genetic effects, leading to the loss of important genetic information in the predictive models.

[0004] Furthermore, in the conservation and breeding decisions of endangered varieties, the cost of each selection decision is extremely high—the erroneous elimination of individuals carrying valuable alleles may lead to irreversible loss of genetic diversity. Therefore, breeding decision-makers not only need to obtain the breeding value prediction results of individuals, but also need to understand the reliability and uncertainty of each prediction value in order to take more prudent strategies in the selection process. However, most existing genome-wide selection methods only output point estimates and lack the ability to quantify the prediction uncertainty on an individual basis.

[0005] In recent years, deep learning methods have shown potential in the field of genome prediction. Existing studies have attempted to apply convolutional neural networks (CNNs), recurrent neural networks (RNNs), or multilayer perceptrons (MLPs) to predict genome selection breeding values. However, these existing methods typically input all SNP sites directly into the network as one-dimensional sequences, failing to fully utilize the functional annotation information of SNPs and the hierarchical structure of the genome, resulting in low model learning efficiency. Simply using convolution or attention mechanisms cannot effectively capture local linkage disequilibrium patterns and long-range epistatic interactions simultaneously. Effective data augmentation and transfer learning strategies have not been designed for small population scenarios, making the model prone to overfitting when the reference population is small. Furthermore, the lack of a mechanism for quantifying prediction uncertainty makes it impossible to provide reliability assessment information for conservation breeding decisions.

[0006] Therefore, there is an urgent need to develop a whole-genome selection breeding value prediction method that can make full use of prior knowledge from genome functional annotation, effectively model epistatic interaction effects, be applicable to small population scenarios, and have the ability to quantify prediction uncertainty, so as to meet the actual needs of conservation and breeding of endangered local livestock and poultry genetic resources. Summary of the Invention

[0007] The purpose of this invention is to address the technical bottlenecks faced by existing whole-genome selection technologies in the conservation and breeding of endangered local livestock and poultry genetic resources, such as small reference population size, difficulty in capturing epistatic effects, and inability to quantify prediction uncertainties. This invention provides a breeding value prediction method and system based on multi-scale gene feature tensors.

[0008] The specific technical solution is as follows: On the one hand, the present invention provides a method for predicting breeding values ​​based on multi-scale gene feature tensors, the method comprising the following steps: Step S1: Obtain whole-genome SNP marker data of the target population, and perform quality control, missing genotype imputation and standardization on the SNP marker data to obtain a preprocessed genotype matrix.

[0009] Step S2: Obtain the phenotypic observations of the target trait in the target population, and apply a linear mixture model to the phenotypic observations for fixed effects correction to obtain the corrected phenotypic values.

[0010] Step S3: Based on gene function annotation information and linkage disequilibrium structure information, the SNP sites in the preprocessed genotype matrix are divided into multiple functional regions. Local feature extraction and dimensionality reduction are performed on each functional region to construct a multi-scale genome feature tensor.

[0011] Step S4: Construct a convolutional-attention fusion deep neural network model that includes a function-aware embedding layer, a hierarchical convolutional attention encoder, and a regression prediction head.

[0012] Step S5: Using the composite loss function as the optimization objective, the deep neural network model is iteratively trained by combining regularization and early stopping strategies to determine the optimal model parameters.

[0013] Step S6: Input the preprocessed genotype data of the individuals to be predicted into the trained model, and output the breeding value prediction results and prediction uncertainty quantification index through multiple random forward propagations.

[0014] Furthermore, the method for obtaining the preprocessed genotype matrix includes the following specific steps: Step S1.1, Obtain the target group Individuals Genotype detection data at SNP loci across the entire genome constitute the original genotype matrix.

[0015] Step S1.2, perform quality control operations on the original genotype matrix, including: removing SNP sites with a minor allele frequency (MAF) lower than a first preset threshold, where the first preset threshold is MAF < 0.01; removing SNP sites or individuals with a genotype deletion rate higher than a second preset threshold, where the second preset threshold is a deletion rate > 10%; and removing SNP sites that have passed the Hardy-Weinberg equilibrium test. SNP sites with values ​​lower than a third preset threshold, wherein the third preset threshold is... .

[0016] Step S1.3: Perform haplotype inference and missing genotype imputation based on Hidden Markov Model on the genotype data after quality control. Specifically, construct the state space of Hidden Markov Model using the haplotype frequency information of the target population or external reference panel, calculate the posterior genotype probability distribution of each missing site under the given known site conditions using the forward-backward algorithm, and select the genotype with the highest posterior probability as the imputation value.

[0017] Step S1.4: For each SNP locus in the imputed genotype matrix Allele frequency centering and unit variance standardization were performed. The standardization formula is as follows: ;in For the first The individual in the first The original genotype coding value of each SNP locus is 0, 1, or 2; The value is the standardized value; For the first Reference allele frequencies of each SNP locus; after standardization, the dimension is obtained as follows: Preprocessed genotype matrix ,in This represents the number of SNP sites retained after quality control.

[0018] Furthermore, the method for obtaining the corrected phenotypic value includes the following specific steps: Step S2.1, obtain information related to the target group. The phenotypic observations of the target trait corresponding to each individual constitute the phenotypic observation vector. .

[0019] Step S2.2, construct the following linear mixture model to estimate the fixed effects: ;in for The vector of phenotypic observations, Design a matrix for fixed effects. This is a fixed effects vector, including field effects, year-season effects, and gender effects. Design a matrix for random effects. This is a vector of random multigene effects. Let be the residual vector; Assume: ;in The genomic kinship matrix was constructed using the VanRaden method: ;in Design a matrix for the centered random effects. For the first Reference allele frequencies at each locus, For additive inheritance variance, For residual variance, It is an identity matrix.

[0020] Step S2.3: Estimate the variance components using the restricted maximum likelihood method. and And based on the Henderson mixture model equations, the fixed effects estimates are solved. Calculate the corrected phenotypic values: The corrected phenotypic value As a response variable for subsequent deep learning model training.

[0021] Furthermore, the method for constructing multi-scale genome feature tensors includes the following specific steps: Step S3.1: Based on the gene function annotation database, label each SNP site in the preprocessed genotype matrix as at least one of the following functional categories: coding region SNP, cis-regulatory element region SNP, or intergenetic region SNP.

[0022] Step S3.2: Within the same functional category, using the chain imbalance attenuation distance as a constraint, a hierarchical clustering method is employed to group physically adjacent and paired LD coefficients. SNP sites exceeding a preset LD threshold are clustered into the same functional region block, where the preset LD threshold is... Forming L functional area blocks Simultaneously, adaptive granularity adjustment is performed: when the number of SNP sites contained in a single functional region block exceeds a preset upper limit. When the number of SNP sites in a functional region block is lower than a preset lower limit, it is subdivided into sub-blocks based on a recursive binary search method; When , merge with adjacent blocks of the same type.

[0023] Step S3.3, for the first Functional area blocks Principal component analysis (PCA) was performed on the SNP-labeled submatrix within the matrix to extract the cumulative variance explained by the top-ranked SNPs that reached a preset variance threshold. The number of principal components, wherein the preset variance threshold is such that the cumulative variance explained is not less than 90%, and the number of principal components... The method for determining it is as follows: ;in For the first The eigenvalues ​​corresponding to each principal component For the first The number of SNP sites within each functional region block is used to obtain the local feature vector of that block. .

[0024] Step S3.4: Organize and arrange the local feature vectors of each functional region block according to their functional category labels and physical locations on the chromosome to construct a multi-scale genome feature tensor. The feature tensor simultaneously encodes the functional annotation hierarchy information of SNP sites, local linkage disequilibrium structure information, and chromosome spatial location information.

[0025] Furthermore, the constructed convolutional-attention fusion deep neural network model includes: a function-aware embedding layer, a hierarchical convolutional attention encoder, and a regression prediction head.

[0026] Furthermore, the function-aware embedding layer receives the multi-scale genomic feature tensor. For those belonging to different functional categories The Local feature vectors of each functional area block Through the independent linear projection matrix corresponding to this function category Mapping to a unified dimension The latent embedding space, and superimposed with learnable functional category encodings. With position encoding Generate functionally aware genome embedding sequences : ; among which the projection matrices of each functional category Parameters are learned independently, and the embedding dimension d ranges from 128 to 512.

[0027] The position encoding uses sine and cosine position encoding, and the formula is as follows: ; Where pos is the sequential index of the functional region block on the chromosome, and i is the index of the embedding dimension; thus, the embedding sequence is generated. .

[0028] Furthermore, the hierarchical convolutional attention encoder includes A series of cascaded coding layers, The values ​​range from 2 to 8. Each coding layer contains three sub-layers in sequence: a grouped one-dimensional depthwise separable convolutional sub-layer, a multi-head self-attention sub-layer, and a feedforward fully connected sub-layer; each of the three sub-layers is configured with residual connections and layer normalization operations.

[0029] The grouped one-dimensional depthwise separable convolutional sublayers group the embedded sequences according to functional categories. Each group independently performs depthwise separable convolution operations. The depthwise separable convolution includes a depthwise convolution part and a pointwise convolution part, wherein the kernel size of the depthwise convolution part is... The sequence length is selected from 3 to 11, with a step size of 1. Causal padding is used to maintain the sequence length. The calculation process is as follows: The depthwise convolution output is then processed point-by-point. Convolution is used for cross-channel information fusion; The multi-head self-attention sublayer is for all The embedding of each functional region block performs global multi-head attention computation, with the number of attention heads... Four to 16 attention heads are selected, with each attention head having a feature dimension of 32 to 128. The multi-head self-attention sub-layer is used to model long-range epistatic nonlinear interaction effects across functional categories and across chromosome regions.

[0030] The feedforward fully connected sublayer consists of two fully connected networks connected by a GELU activation function, with a hidden layer dimension of [missing information]. : , and These are the linear transformation weight matrices, and This is the bias vector.

[0031] Furthermore, the regression prediction head performs a global average pooling operation on the feature sequence finally output by the hierarchical convolutional attention encoder to obtain a fixed-length global genome representation vector, which is then passed through two fully connected layers to finally output a scalar prediction value as the genome estimated breeding value GEBV.

[0032] Furthermore, the calculation of breeding value prediction results and prediction uncertainty quantification indicators includes: inputting the preprocessed genotype data of the individuals to be predicted into the trained model, using the Monte Carlo Dropout inference method, maintaining the activation state of the Dropout layer during the inference phase, and executing... Second random forward propagation, ;by The mean of the prediction results is used as the final estimated breeding value of the individual's genome. ;by The variance of the prediction results is used as a quantitative indicator of prediction uncertainty. ;in, For the completed deep neural network model, is the random mask generated by the Dropout layer during the t-th forward propagation.

[0033] Calculate the prediction reliability score for each individual based on the prediction uncertainty variance: ;in The additive genetic variance estimate is used for the target trait; the prediction reliability score is used to guide selection decisions in breeding and conservation, and individuals with prediction reliability below a preset threshold are marked as individuals requiring careful decision-making.

[0034] On the other hand, the present invention provides a breeding value prediction system based on multi-scale gene feature tensors. The system includes: a genotype data preprocessing module, a phenotypic data correction module, a hierarchical genome feature construction module, a convolution-attention fusion deep prediction module, a model training and optimization module, and a breeding value prediction output module.

[0035] The genotype data preprocessing module is configured to acquire whole-genome SNP marker data of the target population, perform quality control operations including allele frequency filtering, deletion rate filtering, and Hardy-Weinberg balance test, and complete missing value imputation and standardization through a genotype imputation algorithm based on a hidden Markov model, outputting a preprocessed genotype matrix.

[0036] The phenotypic data correction module is configured to acquire phenotypic observations of the target trait, construct a linear mixture model, estimate variance components and fixed effect parameters using the restricted maximum likelihood method, and output corrected phenotypic values ​​after removing the influence of systematic environmental factors.

[0037] The hierarchical genome feature construction module is configured to divide SNP sites into functional categories such as coding regions, cis-regulatory element regions, and intergenic regions based on gene function annotation information. Within each functional category, it performs adaptive granularity functional region block division based on linkage disequilibrium structure, performs principal component analysis on each functional region block to extract local feature vectors, and organizes and constructs multi-scale genome feature tensors.

[0038] The convolutional-attention fusion deep prediction module includes a function-aware embedding layer, a hierarchical convolutional attention encoder, and a regression prediction head. The function-aware embedding layer maps the local feature vectors of different function categories to a unified latent embedding space through independent linear projection matrices for each function category and superimposes learnable function category encoding and sine and cosine position encoding. The hierarchical convolutional attention encoder includes multiple cascaded encoding layers. Each encoding layer sequentially includes a grouped one-dimensional depthwise separable convolutional sublayer, a multi-head self-attention sublayer, and a feedforward fully connected sublayer. Each sublayer is configured with residual connections and layer normalization. The regression prediction head performs global average pooling on the encoder output and outputs the genome-estimated breeding value through a fully connected layer.

[0039] The model training and optimization module is configured to use a weighted combination of breeding value prediction mean squared error loss, L2 regularization loss, and auxiliary supervision loss based on known causal locus effects as a composite loss function. It combines Dropout regularization, cosine annealing learning rate scheduling, and early stopping strategy to iteratively train the convolution-attention fusion deep prediction module. It also includes a data augmentation submodule based on genetic simulation recombination process and a cross-population transfer learning submodule.

[0040] The breeding value prediction output module is configured to input the preprocessed genotype data of the individuals to be predicted into the trained convolutional-attention fusion deep prediction module. It performs multiple random forward propagations through Monte Carlo Dropout inference and outputs the mean of the predicted breeding value, the variance of the prediction uncertainty, and the prediction reliability score. It supports a multi-trait joint prediction mode, which realizes the synchronous prediction of multi-trait breeding values ​​through a multi-task regression head and a cross-attention layer between traits.

[0041] Compared with the prior art, the beneficial effects of this invention are: This invention constructs a multi-scale genome feature tensor by fusing prior knowledge of gene function annotation with linkage disequilibrium structural information. It utilizes a convolutional-attention fusion deep neural network to simultaneously model local linkage disequilibrium patterns and long-range epistatic interaction effects. Furthermore, it combines data augmentation strategies based on genetic simulation and cross-population transfer learning strategies to address the problem of insufficient training samples in small populations. Additionally, it introduces a Monte Carlo Dropout inference mechanism to quantify prediction uncertainty on an individual-by-individual basis. This provides highly accurate breeding value prediction results and reliability assessment information for conservation and breeding decisions of endangered local livestock and poultry breeds. Attached Figure Description

[0042] Figure 1 This is a flowchart of the breeding value prediction method based on multi-scale gene feature tensors of the present invention; Figure 2 This is a schematic diagram of the breeding value prediction system based on multi-scale gene feature tensors of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of this invention, not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0044] Example 1 This embodiment uses a breeding value prediction method based on multi-scale gene feature tensors as an example for detailed explanation.

[0045] To facilitate understanding, the following example illustrates the application scenario of predicting the growth rate trait (daily weight gain) of a certain endangered local pig breed (such as Tibetan pig) through genome-wide selection breeding.

[0046] Assuming the target population consists of 300 Tibetan pigs collected from a Tibetan pig breeding farm, whole-genome genotyping was performed using a 50K pig SNP chip, initially obtaining genotypic data for approximately 50,000 SNP marker loci; at the same time, the daily weight gain phenotypic data of each individual under standardized feeding conditions were recorded.

[0047] The core challenges of this application scenario are: the Tibetan pig population is extremely small (only 300 heads), far smaller than the reference population size of several thousand heads commonly used for whole-genome selection of commercial pig breeds; Tibetan pigs have accumulated a large number of epistatic interactions related to environmental adaptation during long-term high-altitude adaptive evolution; as an endangered breed, the cost of each selection decision is extremely high, requiring a quantitative assessment of the reliability of the prediction results.

[0048] The method provided in this embodiment systematically addresses the aforementioned technical challenges through techniques such as multi-scale genomic feature construction, convolutional-attention fusion modeling, data augmentation and transfer learning, and prediction uncertainty quantification.

[0049] like Figure 1 The diagram shows a flowchart of the breeding value prediction method based on multi-scale gene feature tensors of the present invention. The method includes the following steps: Step S1: Obtain whole-genome SNP marker data of the target population, and perform quality control, missing genotype imputation and standardization on the SNP marker data to obtain a preprocessed genotype matrix.

[0050] The method for obtaining the preprocessed genotype matrix includes the following specific steps: Step S1.1, Obtain the target group Individuals Genotype detection data at SNP loci across the entire genome constitute the original genotype matrix.

[0051] Step S1.2, perform quality control operations on the original genotype matrix, including: removing SNP sites with a minor allele frequency (MAF) lower than a first preset threshold, where the first preset threshold is MAF < 0.01; removing SNP sites or individuals with a genotype deletion rate higher than a second preset threshold, where the second preset threshold is a deletion rate > 10%; and removing SNP sites that have passed the Hardy-Weinberg equilibrium test. SNP sites with values ​​lower than a third preset threshold, wherein the third preset threshold is... .

[0052] Step S1.3: Perform haplotype inference and missing genotype imputation based on Hidden Markov Model on the genotype data after quality control. Specifically, construct the state space of Hidden Markov Model using the haplotype frequency information of the target population or external reference panel, calculate the posterior genotype probability distribution of each missing site under the given known site conditions using the forward-backward algorithm, and select the genotype with the highest posterior probability as the imputation value.

[0053] Step S1.4: For each SNP locus in the imputed genotype matrix Allele frequency centering and unit variance standardization were performed. The standardization formula is as follows: ;in For the first The individual in the first The original genotype coding value of each SNP locus is 0, 1, or 2; The value is the standardized value; For the first Reference allele frequencies of each SNP locus; after standardization, the dimension is obtained as follows: Preprocessed genotype matrix ,in This represents the number of SNP sites retained after quality control.

[0054] Taking the Tibetan pig application scenario mentioned above as an example, the original genotype matrix is ​​a 300×50,000 matrix. In the quality control step S1.2, it is assumed that approximately 5,000 low-frequency loci with MAF < 0.01 are removed, approximately 2,000 loci with a genotype deletion rate exceeding 10% are removed, and approximately 500 loci that fail the HWE test are removed, leaving approximately 42,500 SNP loci after quality control.

[0055] In step S1.3, missing genotypes are imputed using haplotype frequency information from an external pig reference panel (such as pig genome data from the 1000 Bull Genomes Project). After standardization in step S1.4, a preprocessed genotype matrix X with dimensions 300×42,500 is obtained. Standardization eliminates scale inconsistencies caused by allele frequency differences at different SNP loci, allowing the subsequent deep learning model to assign appropriate weights to each point.

[0056] Step S2: Obtain the phenotypic observations of the target trait in the target population, and apply a linear mixture model to the phenotypic observations for fixed effects correction to obtain the corrected phenotypic values.

[0057] The method for obtaining the corrected phenotypic value includes the following specific steps: Step S2.1, obtain information related to the target group. The phenotypic observations of the target trait corresponding to each individual constitute the phenotypic observation vector. .

[0058] Step S2.2, construct the following linear mixture model to estimate the fixed effects: ;in for The vector of phenotypic observations, Design a matrix for fixed effects. This is a fixed effects vector, including field effects, year-season effects, gender effects, and other known systematic environmental factors. Design a matrix for random effects. This is a vector of random multigene effects. Let be the residual vector; Assume: ;in The genomic kinship matrix was constructed using the VanRaden method: ;in Design a matrix for the centered random effects. For the first Reference allele frequencies at each locus, For additive inheritance variance, For residual variance, It is an identity matrix.

[0059] Step S2.3: Estimate the variance components using the restricted maximum likelihood method. and And based on the Henderson mixture model equations, the fixed effects estimates are solved. Calculate the corrected phenotypic values: The corrected phenotypic value As a response variable for subsequent deep learning model training.

[0060] Step S3: Based on gene function annotation information and linkage disequilibrium structure information, the SNP sites in the preprocessed genotype matrix are divided into multiple functional regions. Local feature extraction and dimensionality reduction are performed on each functional region to construct a multi-scale genome feature tensor.

[0061] The method for constructing multi-scale genome feature tensors includes the following specific steps: Step S3.1: Based on the gene function annotation database, label each SNP site in the preprocessed genotype matrix as at least one of the following functional categories: coding region SNP (located in the exon region, affecting the protein coding sequence), cis-regulatory element region SNP (located in the regulatory regions such as promoters, enhancers, and UTRs), or intergene region SNP (located in the intergene region).

[0062] Step S3.2: Within the same functional category, using the chain imbalance attenuation distance as a constraint, a hierarchical clustering method is employed to group physically adjacent and paired LD coefficients. SNP sites exceeding a preset LD threshold are clustered into the same functional region block, where the preset LD threshold is... Forming L functional area blocks Simultaneously, adaptive granularity adjustment is performed: when the number of SNP sites contained in a single functional region block exceeds a preset upper limit. When the number of SNP sites in a functional region block is lower than a preset lower limit, it is subdivided into sub-blocks based on a recursive binary search method; When , merge with adjacent blocks of the same type.

[0063] Step S3.3, for the first Functional area blocks Principal component analysis (PCA) was performed on the SNP-labeled submatrix within the matrix to extract the cumulative variance explained by the top-ranked SNPs that reached a preset variance threshold. The number of principal components, wherein the preset variance threshold is such that the cumulative variance explained is not less than 90%, and the number of principal components... The method for determining it is as follows: ;in For the first The eigenvalues ​​corresponding to each principal component For the first The number of SNP sites within each functional region block is used to obtain the local feature vector of that block. .

[0064] Step S3.4: Organize and arrange the local feature vectors of each functional region block according to their functional category labels and physical locations on the chromosome to construct a multi-scale genome feature tensor. The feature tensor simultaneously encodes the functional annotation hierarchy information of SNP sites, local linkage disequilibrium structure information, and chromosome spatial location information.

[0065] Continuing with the Tibetan pig application scenario mentioned above, in step S3, 42,500 quality-controlled SNP sites are functionally labeled using gene function annotation databases such as Ensembl. It is assumed that approximately 3,500 of these are located in coding regions (exon regions), approximately 8,000 are located in cis-regulatory element regions (promoters, enhancers, UTRs, etc.), and the remaining approximately 31,000 are located in intergenic regions. Taking coding region SNPs as an example, hierarchical clustering is performed within the same functional category using an LD coefficient r² ≥ 0.3 as a threshold. For instance, there are 15 coding region SNPs near the IGF2 gene on pig chromosome 6, which is related to muscle growth and development. Because their LD coefficients are all high (r² ≥ 0.3), they are clustered into the same functional region block.

[0066] The upper limit of the number of SNP sites within a functional region block was set to 50, and the lower limit to 5. Through adaptive granularity adjustment, the 42,500 SNP sites were finally divided into approximately 800 functional region blocks.

[0067] PCA dimensionality reduction was performed on each functional region block. Principal components were selected based on a cumulative variance explanation rate of at least 90%. For example, the IGF2 region block containing 15 SNP sites achieved a 90% cumulative variance explanation rate by retaining only the first four principal components after PCA. The final constructed multi-scale genome feature tensor F compressed the 42,500-dimensional high-dimensional genotype data into a combination of local feature vectors from approximately 800 functional regions, significantly reducing data dimensionality while preserving key genetic information structures.

[0068] Step S4: Construct a convolutional-attention fusion deep neural network model that includes a function-aware embedding layer, a hierarchical convolutional attention encoder, and a regression prediction head.

[0069] The constructed convolutional-attention fusion deep neural network model includes: a function-aware embedding layer, a hierarchical convolutional attention encoder, and a regression prediction head.

[0070] The function-aware embedding layer receives the multi-scale genomic feature tensor. For those belonging to different functional categories The Local feature vectors of each functional area block Through the independent linear projection matrix corresponding to this function category Mapping to a unified dimension The latent embedding space, and superimposed with learnable functional category encodings. With position encoding Generate functionally aware genome embedding sequences : ; among which the projection matrices of each functional category Parameters are learned independently, and the embedding dimension d ranges from 128 to 512. The position encoding uses sine and cosine position encoding, and the formula is as follows: ; Where pos is the sequential index of the functional region block on the chromosome, and i is the index of the embedding dimension; thus, the embedding sequence is generated. .

[0071] The hierarchical convolutional attention encoder includes A series of cascaded coding layers, The values ​​range from 2 to 8. Each coding layer contains three sub-layers in sequence: a grouped one-dimensional depthwise separable convolutional sub-layer, a multi-head self-attention sub-layer, and a feedforward fully connected sub-layer; each of the three sub-layers is configured with residual connections and layer normalization operations. The grouped one-dimensional depthwise separable convolutional sublayers group the embedded sequences according to functional categories. Each group independently performs depthwise separable convolution operations. The depthwise separable convolution includes a depthwise convolution part and a pointwise convolution part, wherein the kernel size of the depthwise convolution part is... The sequence length is selected from 3 to 11, with a step size of 1. Causal padding is used to maintain the sequence length. The calculation process is as follows: The depthwise convolution output is then processed point-by-point. Convolution is used for cross-channel information fusion; The multi-head self-attention sublayer is for all The embedding of each functional region block performs global multi-head attention computation, with the number of attention heads... Choose 4 to 16 attention heads, with each attention head having a feature dimension of 32 to 128; the calculation process is as follows: in These are query, key, and value matrices, respectively. For each attention head, the feature dimensions, The output projection matrix is ​​used; the multi-head self-attention sublayer is used to model long-range epistatic nonlinear interaction effects across functional categories and across chromosome regions; The feedforward fully connected sublayer consists of two fully connected networks connected by a GELU activation function, with a hidden layer dimension of [missing information]. : , and These are the linear transformation weight matrices, and This is the bias vector.

[0072] The regression prediction head performs global average pooling on the feature sequence finally output by the hierarchical convolutional attention encoder to obtain a fixed-length global genome representation vector. This vector is then passed through two fully connected layers (connected by a ReLU activation function and subjected to Dropout regularization) to finally output a scalar prediction value as the genome estimated breeding value GEBV.

[0073] Step S5: Using the composite loss function as the optimization objective, the deep neural network model is iteratively trained by combining regularization and early stopping strategies to determine the optimal model parameters.

[0074] Step S5 specifically includes the following steps: Step S5.1: Randomly divide the preprocessed genotype matrix and the corresponding corrected phenotypic values ​​into training sets according to a preset ratio. and verification set The ratio of division is 8:2 or a five-fold cross-validation method is adopted.

[0075] Step S5.2, perform end-to-end training with the following composite loss function as the optimization objective: The definitions of each loss term are as follows: Breeding value prediction mean square error loss : ;in, The model outputs its predicted breeding value for the i-th individual. is the corresponding corrected phenotypic value, and n is the number of training set samples.

[0076] L2 regularization loss : ;in, These are the learnable parameters for each layer of the model. The L2 regularization coefficient has a range of values. .

[0077] Auxiliary monitoring loss An auxiliary decoding branch is derived from the intermediate encoding layer of the hierarchical convolutional attention encoder. This auxiliary decoding branch comprises two fully connected layers. The input is the functional region block-level embedding features output from the intermediate encoding layer, and the output is the predicted effect size corresponding to each functional region block. The auxiliary supervision loss is calculated using pre-obtained estimates of the effect size of known causal loci or candidate gene regions as auxiliary supervision labels. Where S is the set of functional regions corresponding to known causal loci. To assist in the prediction effect size of the decoding branch, Labels for the magnitude of prior effects. The auxiliary loss balancing coefficient has a value range of 0.01 to 0.5.

[0078] Step S5.3: Update parameters using the Adam optimizer, with the initial learning rate set to... The learning rate scheduling strategy is cosine annealing; Dropout regularization is applied at the output of the self-attention sub-layer and feedforward sub-layer of each coding layer, with the Dropout probability ranging from 0.1 to 0.3.

[0079] Step S5.4, with the verification set The mean squared error is used as the monitoring metric for implementing an early stopping strategy when the validation set loss is continuous. Training should be terminated if no improvement is observed after a certain number of training rounds. Select parameters from 20 to 50, and save the model parameters at the moment when the validation set loss is the lowest as the optimal model.

[0080] Step S5 also includes a data augmentation step based on genetic simulation to address the problem of insufficient training samples for endangered small populations, specifically: Step S5a.1: Randomly select two individuals from the existing genotype data of the target population as virtual parent pairs. Based on the allele frequency spectrum and linkage disequilibrium structure information of the target population, simulate the meiotic recombination process to generate synthetic offspring genotype samples. Specifically, along each chromosome, exchange sites are randomly generated according to the Poisson process with the average recombination rate as a parameter. At the exchange sites, the two parent haplotypes are cross-recombined to generate new offspring haplotype combinations.

[0081] Step S5a.2: Perform forward inference on the synthetic progeny genotype samples using the currently trained model to obtain the predicted breeding value. As pseudo-labels; add noise perturbation based on prediction uncertainty to the pseudo-labels: ,in This is the estimated variance of the uncertainty in the current model's prediction of this synthetic sample.

[0082] Step S5a.3: Add the synthesized progeny genotype samples and their noisy pseudo-labels to the training set. The number of synthetic samples is 0.5 to 2 times that of the original training set, and they participate in subsequent model training iterations; synthetic samples are regenerated and pseudo-labels are updated every preset training round.

[0083] Step S5 also includes a cross-population transfer learning step to transfer the general genomic representation knowledge learned from a large-scale reference population to an endangered small population, specifically: Step S5b.1, Pre-training stage: Using genotype and phenotypic data from a large-scale reference population (with a population size of no less than 2000 individuals), the convolutional-attention fusion deep neural network model is fully trained according to the process of steps S1 to S5 to learn a general genomic feature representation. The large-scale reference population and the target small population are different varieties of the same species or populations with different genetic backgrounds.

[0084] Step S5b.2, Parameter Freezing and Fine-tuning Stage: Freeze all parameters of the function-aware embedding layer and the bottom layer of the hierarchical convolutional attention encoder. The parameters of each encoding layer are fine-tuned using only target small group data (group size can be as low as 100 individuals) to fine-tune the parameters of the top encoding layer and the regression prediction head in the hierarchical convolutional attention encoder; the learning rate in the fine-tuning stage is set to 1 / 10 to 1 / 100 of the learning rate in the pre-training stage, and the number of training rounds does not exceed 1 / 5 of the pre-training stage.

[0085] Three hundred Tibetan pigs were randomly divided into a training set (240 pigs) and a validation set (60 pigs) at an 8:2 ratio. During the data augmentation phase, parental pairs were randomly selected from the 240 training individuals for simulated recombination. For example, individuals A and B were randomly selected as virtual parents. Recombination sites were generated along the 18 autosomes of the pig according to the Poisson process at a recombination rate of one recombination per 100 cM on average. At each recombination site, the haplotypes of the two parents were cross-recombined to generate synthetic offspring genotypes.

[0086] Approximately 240–480 synthetic samples (1–2 times the original training set) are generated in this manner, using the breeding values ​​predicted by the current model as pseudo-labels and adding noise perturbations based on prediction uncertainty.

[0087] In cross-population transfer learning, the model was first pre-trained using genotypic and daily weight gain phenotypic data from 3,000 individuals of a Large White pig breed to learn general genetic features related to growth rate in the pig genome. After pre-training, the parameters of the functional-aware embedding layer and the bottom two coding layers were frozen, and the top coding layer and regression prediction head were fine-tuned using only training data from 240 Tibetan pigs. Through transfer learning, the model was able to transfer growth-related gene effect patterns learned in Large White pigs (such as the effect characteristics of known growth genes like MC4R and IGF1) to the Tibetan pig population, compensating for the insufficient size of the Tibetan pig reference population.

[0088] Step S6: Input the preprocessed genotype data of the individuals to be predicted into the trained model, and output the breeding value prediction results and prediction uncertainty quantification index through multiple random forward propagations.

[0089] The calculation of breeding value prediction results and prediction uncertainty quantification indicators includes: inputting preprocessed genotype data of the individuals to be predicted into the trained model, using Monte Carlo Dropout inference, maintaining the activation state of the Dropout layer during the inference phase, and executing... Second random forward propagation, ;by The mean of the prediction results is used as the final estimated breeding value of the individual's genome. ;by The variance of the prediction results is used as a quantitative indicator of prediction uncertainty. ;in, For the completed deep neural network model, is the random mask generated by the Dropout layer during the t-th forward propagation.

[0090] Calculate the prediction reliability score for each individual based on the prediction uncertainty variance: ;in The additive genetic variance estimate is used for the target trait; the prediction reliability score is used to guide selection decisions in breeding and conservation, and individuals with prediction reliability below a preset threshold are marked as individuals requiring careful decision-making.

[0091] For a candidate Tibetan pig boar to be evaluated, its preprocessed genotype data (1×42,500-dimensional vector) is input into a trained model, with the number of Monte Carlo Dropout inferences T=50. In each forward propagation, the model output varies slightly due to the different random masks in the Dropout layer. Assuming the 50 prediction results are 0.85, 0.82, 0.88, ..., the mean of 0.856 is taken as the final genome estimated breeding value (GEBV) for this individual, indicating that this individual has a genetic advantage in the daily weight gain trait that is 0.856 standard deviations higher than the population average. The variance of the 50 prediction results is 0.012, reflecting the uncertainty of the model regarding this prediction result.

[0092] Assuming the additive genetic variance estimate of the target trait is 0.15, the prediction reliability score (REL) for this individual is 1 - 0.012 / 0.15 = 0.92, indicating that the prediction result has high reliability. If the prediction uncertainty variance of another candidate individual is 0.08, then its REL is 1 - 0.08 / 0.15 = 0.47, which is lower than the preset threshold (e.g., 0.5). This individual will be marked as an individual requiring careful decision-making. It is recommended that breeding decision-makers combine other information (such as pedigree records, progeny determination, etc.) when selecting or culling this individual to avoid the wrong culling of individuals carrying valuable alleles due to inaccurate predictions.

[0093] The method also supports a multi-trait joint breeding value prediction model, specifically: The regression prediction head is replaced with a multi-task regression head structure containing P parallel output branches, each output branch corresponding to the breeding value prediction of a target trait, where P is the number of target traits; After the top coding layer of the hierarchical convolutional attention encoder, a trait cross attention layer is introduced. The trait cross attention layer performs cross attention calculation using the feature representation of each trait as a query and key-value pair to model the genetic covariance structure between different traits. Joint training is performed using the following multi-task learning loss function: ;in The mean squared error loss for the p-th trait is... Let the task weight be the p-th trait, and the task weight be determined as follows: ;in Let p be the estimate of the heritability of the p-th trait. denoted as the economic weight coefficient of the p-th trait.

[0094] Example 2 This embodiment provides a breeding value prediction system based on multi-scale gene feature tensors, used to execute the breeding value prediction method described in Embodiment 1.

[0095] The system adopts a modular architecture design, with each functional module transmitting data and working collaboratively through standardized data interfaces. It can be deployed on high-performance computing servers to meet the batch prediction needs of large-scale populations, or on workstations equipped with GPU accelerator cards to meet the daily breeding value evaluation needs of small and medium-sized breeding farms.

[0096] like Figure 2 The diagram shown illustrates the components of the breeding value prediction system based on multi-scale gene feature tensors of the present invention. The system includes: a genotype data preprocessing module, a phenotypic data correction module, a hierarchical genome feature construction module, a convolution-attention fusion deep prediction module, a model training and optimization module, and a breeding value prediction output module.

[0097] The genotype data preprocessing module is configured to acquire whole-genome SNP marker data of the target population, perform quality control operations including allele frequency filtering, deletion rate filtering, and Hardy-Weinberg balance test, and complete missing value imputation and standardization through a genotype imputation algorithm based on a hidden Markov model, outputting a preprocessed genotype matrix.

[0098] The phenotypic data correction module is configured to acquire phenotypic observations of the target trait, construct a linear mixture model, estimate variance components and fixed effect parameters using the restricted maximum likelihood method, and output corrected phenotypic values ​​after removing the influence of systematic environmental factors.

[0099] The hierarchical genome feature construction module is configured to divide SNP sites into functional categories such as coding regions, cis-regulatory element regions, and intergenic regions based on gene function annotation information. Within each functional category, it performs adaptive granularity functional region block division based on linkage disequilibrium structure, performs principal component analysis on each functional region block to extract local feature vectors, and organizes and constructs multi-scale genome feature tensors.

[0100] The convolutional-attention fusion deep prediction module includes a function-aware embedding layer, a hierarchical convolutional attention encoder, and a regression prediction head. The function-aware embedding layer maps the local feature vectors of different function categories to a unified latent embedding space through independent linear projection matrices for each function category and superimposes learnable function category encoding and sine and cosine position encoding. The hierarchical convolutional attention encoder includes multiple cascaded encoding layers. Each encoding layer sequentially includes a grouped one-dimensional depthwise separable convolutional sublayer, a multi-head self-attention sublayer, and a feedforward fully connected sublayer. Each sublayer is configured with residual connections and layer normalization. The regression prediction head performs global average pooling on the encoder output and outputs the genome-estimated breeding value through a fully connected layer.

[0101] The model training and optimization module is configured to use a weighted combination of breeding value prediction mean squared error loss, L2 regularization loss, and auxiliary supervision loss based on known causal locus effects as a composite loss function. It combines Dropout regularization, cosine annealing learning rate scheduling, and early stopping strategy to iteratively train the convolution-attention fusion deep prediction module. It also includes a data augmentation submodule based on genetic simulation recombination process and a cross-population transfer learning submodule.

[0102] The breeding value prediction output module is configured to input the preprocessed genotype data of the individuals to be predicted into the trained convolutional-attention fusion deep prediction module. It performs multiple random forward propagations through Monte Carlo Dropout inference and outputs the mean of the predicted breeding value, the variance of the prediction uncertainty, and the prediction reliability score. It supports a multi-trait joint prediction mode, which realizes the synchronous prediction of multi-trait breeding values ​​through a multi-task regression head and a cross-attention layer between traits.

[0103] Taking the Tibetan pig conservation and breeding application scenario as an example, the collaborative workflow of the above modules is explained: The genotype data preprocessing module receives 50K chip genotyping data from 300 Tibetan pigs in the breeding farm, and outputs a 300×42,500 preprocessed genotype matrix after quality control; the phenotypic data correction module simultaneously processes the daily weight gain phenotypic data, removes field effects and seasonal effects, and outputs a corrected phenotypic value vector.

[0104] The hierarchical genome feature construction module uses the Ensembl functional annotation database to divide 42,500 SNPs into three major functional categories: coding regions, cis-regulatory element regions, and intergenic regions. Within each category, approximately 800 functional region blocks are formed based on LD structure clustering. After PCA dimensionality reduction, a multi-scale genome feature tensor is output.

[0105] The model training and optimization module takes the aforementioned feature tensors and corrected phenotypic values ​​as inputs to drive the convolution-attention fusion deep prediction module for iterative training. The data augmentation submodule generates synthetic samples to expand the training set by simulating meiosis and recombination, while the cross-population transfer learning submodule uses the pre-trained model of the Great White pig for parameter transfer and fine-tuning.

[0106] After training, the breeding value prediction output module performs batch predictions on a new batch of candidate breeding pigs to be evaluated in the breeding farm. It uses Monte Carlo Dropout to infer and output the predicted breeding value, uncertainty variance, and reliability score for each individual, and presents them to breeding decision-makers in the form of a structured report.

[0107] Breeding decision-makers can rank and select candidate individuals based on predicted breeding values, while also referring to reliability scores to identify individuals whose predictions are uncertain. For these individuals, conservative strategies can be adopted (such as delaying elimination or waiting for progeny testing results), thereby maximizing the protection of the population's genetic diversity while improving genetic progress.

[0108] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A breeding value prediction method based on multi-scale gene feature tensors, characterized in that, The method includes the following steps: Step S1: Obtain whole-genome SNP marker data of the target population, and perform quality control, missing genotype imputation and standardization on the SNP marker data to obtain a preprocessed genotype matrix; Step S2: Obtain the phenotypic observations of the target trait in the target population, and apply a linear mixture model to the phenotypic observations for fixed effects correction to obtain the corrected phenotypic values; Step S3: Based on gene function annotation information and linkage disequilibrium structure information, the SNP sites in the preprocessed genotype matrix are divided into multiple functional regions. Local feature extraction and dimensionality reduction are performed on each functional region to construct a multi-scale genome feature tensor. Step S4: Construct a convolutional-attention fusion deep neural network model that includes a function-aware embedding layer, a hierarchical convolutional attention encoder, and a regression prediction head; Step S5: Using the composite loss function as the optimization objective, the deep neural network model is iteratively trained by combining regularization and early stopping strategies to determine the optimal model parameters. Step S6: Input the preprocessed genotype data of the individuals to be predicted into the trained model, and output the breeding value prediction results and prediction uncertainty quantification index through multiple random forward propagations.

2. The method according to claim 1, characterized in that, The method for obtaining the preprocessed genotype matrix includes the following specific steps: Step S1.1, Obtain the target group Individuals Genotype detection data at SNP loci across the entire genome constitute the original genotype matrix; Step S1.2, perform quality control operations on the original genotype matrix, including: removing SNP sites with a minor allele frequency (MAF) lower than a first preset threshold, where the first preset threshold is MAF < 0.01; removing SNP sites or individuals with a genotype deletion rate higher than a second preset threshold, where the second preset threshold is a deletion rate > 10%; and removing SNP sites that have passed the Hardy-Weinberg equilibrium test. SNP sites with values ​​lower than a third preset threshold, wherein the third preset threshold is... ; Step S1.3: Perform haplotype inference and missing genotype imputation based on hidden Markov model on the genotype data after quality control. Specifically, construct the state space of the hidden Markov model using the haplotype frequency information of the target population or external reference panel, calculate the posterior genotype probability distribution of each missing site under the given known site conditions using the forward-backward algorithm, and select the genotype with the highest posterior probability as the imputation value. Step S1.4: For each SNP locus in the imputed genotype matrix Allele frequency centering and unit variance standardization were performed. The standardization formula is as follows: ;in For the first The individual in the first The original genotype coding value of each SNP locus is 0, 1, or 2; The value is the standardized value; For the first Reference allele frequencies of each SNP locus; after standardization, the dimension is obtained as follows: Preprocessed genotype matrix ,in This represents the number of SNP sites retained after quality control.

3. The method according to claim 2, characterized in that, The method for obtaining the corrected phenotypic value includes the following specific steps: Step S2.1, obtain information related to the target group. The phenotypic observations of the target trait corresponding to each individual constitute the phenotypic observation vector. ; Step S2.2, construct the following linear mixture model to estimate the fixed effects: ;in for The vector of phenotypic observations, Design a matrix for fixed effects. This is a fixed effects vector, including field effects, year-season effects, and gender effects. Design a matrix for random effects. This is a vector of random multigene effects. Let be the residual vector; Assume: ;in The genomic kinship matrix was constructed using the VanRaden method: ;in Design a matrix for the centered random effects. For the first Reference allele frequencies at each locus, For additive inheritance variance, For residual variance, It is the identity matrix; Step S2.3: Estimate the variance components using the restricted maximum likelihood method. and And based on the Henderson mixture model equations, the fixed effects estimates are solved. Calculate the corrected phenotypic values: The corrected phenotypic value As a response variable for subsequent deep learning model training.

4. The method according to claim 1, characterized in that, The method for constructing multi-scale genome feature tensors includes the following specific steps: Step S3.1: Based on the gene function annotation database, label each SNP site in the preprocessed genotype matrix as at least one of the following functional categories: coding region SNP, cis-regulatory element region SNP, or intergene region SNP. Step S3.2: Within the same functional category, using the chain imbalance attenuation distance as a constraint, a hierarchical clustering method is employed to group physically adjacent and paired LD coefficients. SNP sites exceeding a preset LD threshold are clustered into the same functional region block, where the preset LD threshold is... Forming L functional area blocks Simultaneously, adaptive granularity adjustment is performed: when the number of SNP sites contained in a single functional region block exceeds a preset upper limit. When the number of SNP sites in a functional region block is lower than a preset lower limit, it is subdivided into sub-blocks based on a recursive binary search method; When merging with adjacent blocks of the same type; Step S3.3, for the first Functional area blocks Principal component analysis (PCA) was performed on the SNP-labeled submatrix within the matrix to extract the cumulative variance explained by the top-ranked SNPs that reached a preset variance threshold. The number of principal components, wherein the preset variance threshold is such that the cumulative variance explained is not less than 90%, and the number of principal components... The method for determining it is as follows: ;in For the first The eigenvalues ​​corresponding to each principal component For the first The number of SNP sites within each functional region block is used to obtain the local feature vector of that block. ; Step S3.4: Organize and arrange the local feature vectors of each functional region block according to their functional category labels and physical locations on the chromosome to construct a multi-scale genome feature tensor. The feature tensor simultaneously encodes the functional annotation hierarchy information of SNP sites, local linkage disequilibrium structure information, and chromosome spatial location information.

5. The method according to claim 4, characterized in that, The constructed convolutional-attention fusion deep neural network model includes: a function-aware embedding layer, a hierarchical convolutional attention encoder, and a regression prediction head.

6. The method according to claim 5, characterized in that, The function-aware embedding layer receives the multi-scale genomic feature tensor. For those belonging to different functional categories The Local feature vectors of each functional area block Through the independent linear projection matrix corresponding to this function category Mapping to a unified dimension The potential embedding space; Superimposed learnable function category coding With position encoding Generate functionally aware genome embedding sequences : ; among which the projection matrices of each functional category Parameters are learned independently, and the embedding dimension d ranges from 128 to 512. The position encoding uses sine and cosine position encoding, and the formula is as follows: ; Where pos is the sequential index of the functional region block on the chromosome. An index for the embedding dimension; this generates an embedding sequence. .

7. The method according to claim 6, characterized in that, The hierarchical convolutional attention encoder includes A series of cascaded coding layers, The values ​​range from 2 to 8. Each coding layer contains three sub-layers in sequence: a grouped one-dimensional depthwise separable convolutional sub-layer, a multi-head self-attention sub-layer, and a feedforward fully connected sub-layer; each of the three sub-layers is configured with residual connections and layer normalization operations. The grouped one-dimensional depthwise separable convolutional sublayers group the embedded sequences according to functional categories. Each group independently performs depthwise separable convolution operations. The depthwise separable convolution includes a depthwise convolution part and a pointwise convolution part, wherein the kernel size of the depthwise convolution part is... The sequence length is selected from 3 to 11, with a step size of 1. Causal padding is used to maintain the sequence length. The calculation process is as follows: The depthwise convolution output is then processed point-by-point. Convolution is used for cross-channel information fusion; The multi-head self-attention sublayer is for all The embedding of each functional region block performs global multi-head attention computation, with the number of attention heads... The number of attention heads can be selected from 4 to 16, and the feature dimension of each attention head can be from 32 to 128. The multi-head self-attention sub-layer is used to model long-range epistatic nonlinear interaction effects across functional categories and across chromosome regions. The feedforward fully connected sublayer consists of two fully connected networks connected by a GELU activation function, with a hidden layer dimension of [missing information]. : , and These are the linear transformation weight matrices, and This is the bias vector.

8. The method according to claim 7, characterized in that, The regression prediction head performs global average pooling on the feature sequence finally output by the hierarchical convolutional attention encoder to obtain a fixed-length global genome representation vector. After passing through two fully connected layers, it finally outputs a scalar prediction value as the genome estimated breeding value (GEBV).

9. The method according to claim 8, characterized in that, The calculation of breeding value prediction results and prediction uncertainty quantification indicators includes: inputting preprocessed genotype data of the individuals to be predicted into the trained model, using Monte Carlo Dropout inference, maintaining the activation state of the Dropout layer during the inference phase, and executing... Second random forward propagation, ;by The mean of the prediction results is used as the final estimated breeding value of the individual's genome. ;by The variance of the prediction results is used as a quantitative indicator of prediction uncertainty. ;in, For the completed deep neural network model, The random mask generated by the Dropout layer during the t-th forward propagation; Calculate the prediction reliability score for each individual based on the prediction uncertainty variance: ;in The additive genetic variance estimate is used for the target trait; the prediction reliability score is used to guide selection decisions in breeding and conservation, and individuals with prediction reliability below a preset threshold are marked as individuals requiring careful decision-making.

10. A breeding value prediction system based on multi-scale gene feature tensors, used to execute the method according to any one of claims 1 to 9, characterized in that, The system includes: a genotype data preprocessing module, a phenotypic data correction module, a hierarchical genome feature construction module, a convolution-attention fusion deep prediction module, a model training and optimization module, and a breeding value prediction output module; The genotype data preprocessing module is configured to acquire whole-genome SNP marker data of the target population, perform quality control operations including allele frequency filtering, deletion rate filtering and Hardy-Weinberg balance test, and complete missing value imputation and standardization through a genotype imputation algorithm based on a hidden Markov model, and output a preprocessed genotype matrix. The phenotypic data correction module is configured to acquire phenotypic observations of the target trait, construct a linear mixture model, estimate the variance components and fixed effect parameters using the restricted maximum likelihood method, and output the corrected phenotypic values ​​after removing the influence of systematic environmental factors. The hierarchical genome feature construction module is configured to divide SNP sites into functional categories such as coding regions, cis-regulatory element regions, and intergenic regions based on gene function annotation information. Within each functional category, it performs adaptive granular functional region block division based on linkage disequilibrium structure, performs principal component analysis to extract local feature vectors for each functional region block, and organizes and constructs multi-scale genome feature tensors. The convolutional-attention fusion deep prediction module includes a function-aware embedding layer, a hierarchical convolutional attention encoder, and a regression prediction head. The function-aware embedding layer maps the local feature vectors of different function categories to a unified latent embedding space through independent linear projection matrices for each function category, and superimposes learnable function category encoding and sine / cosine position encoding. The hierarchical convolutional attention encoder includes multiple cascaded encoding layers, each of which sequentially contains a grouped one-dimensional depthwise separable convolutional sublayer, a multi-head self-attention sublayer, and a feedforward fully connected sublayer. Each sublayer is configured with residual connections and layer normalization. The regression prediction head performs global average pooling on the encoder output and outputs the genome-estimated breeding value through a fully connected layer. The model training and optimization module is configured to use a weighted combination of the breeding value prediction mean squared error loss, L2 regularization loss, and auxiliary supervision loss based on known causal locus effects as a composite loss function. It combines Dropout regularization, cosine annealing learning rate scheduling, and early stopping strategy to iteratively train the convolution-attention fusion deep prediction module. It also includes a data augmentation submodule based on genetic simulation recombination process and a cross-population transfer learning submodule. The breeding value prediction output module is configured to input the preprocessed genotype data of the individuals to be predicted into the trained convolutional-attention fusion deep prediction module. It performs multiple random forward propagations through Monte Carlo Dropout inference and outputs the mean of the predicted breeding value, the variance of the prediction uncertainty, and the prediction reliability score. It supports a multi-trait joint prediction mode, which realizes the synchronous prediction of multi-trait breeding values ​​through a multi-task regression head and a cross-attention layer between traits.

Images