Model and method for predicting biological traits by integrating genetic background and variation information

Abstract

The invention discloses a model and method for predicting biological characters by integrating genetic background and variation information, and the model comprises an embedded Siamese network which generates and extracts genetic background and genetic correlation information between samples, and builds an embedded vector for each sample to represent the genetic background information; the genetic variation embedded network is used for extracting complex high-order information and establishing an embedded vector by using a deep learning network and the whole genome variation information of each sample so as to characterize the cumulative effect of genetic variation on characters; and the fusion and character prediction module organically fuses embedded vectors output by the two modules to realize accurate prediction of characters. According to the method, the problem that genetic background information and genome variation information cannot be effectively fused in the prior art can be solved, the nonlinear effect between genes is effectively captured, and the accuracy of genome selection and the genetic architecture analysis capability of complex characters are greatly improved.

Description

Technical Field

[0001] The present invention relates to the fields of biological trait prediction, animal and plant breeding and artificial intelligence, and specifically to a model method for predicting biological traits by integrating genetic background and variation information. Background Art

[0002] Predicting phenotypes based on genotypes is widely used in medicine, animal and plant genetics, and biological research. In animal and plant breeding, the relevant method is often called whole genome selection. Existing prediction methods usually rely on linear models. Commonly used methods such as GBLUP and rrBLUP perform well in predicting some traits, but they cannot capture the nonlinear relationship between genes. These methods show limitations in predictive ability when dealing with complex traits.

[0003] Some deep learning methods are used for genotype-based phenotypic prediction to improve nonlinear prediction capabilities. However, existing deep learning models are not very effective in capturing the complex relationships of genetic variation. Most methods have acceptable prediction accuracy for individual or partial tested traits, but for the remaining traits, the prediction capabilities are often poor and lack sufficient robustness. Existing methods also have shortcomings in the integration of genetic background information. Most methods focus on genomic variation and ignore the role of genetic background, resulting in low prediction accuracy and insufficient model robustness.

[0004] MMNet has the following features:

[0005] 1. Accuracy and robustness of MMNet’s prediction performance:

[0006] The MMNet model performed well on 12 rice agronomic traits (such as yield, protein content, heading date, etc.), especially in terms of prediction ability, which significantly exceeded traditional linear models (such as GBLUP) and other existing deep learning models. Through its efficient prediction ability, MMNet achieved higher accuracy in the prediction of multiple traits.

[0007] 2. MMNet adaptive weight adjustment mechanism:

[0008] MMNet adaptively adjusts the relative amount of information between genetic background information and genomic variation during training. Specifically, the model can adaptively combine genetic background information and site variation information to achieve higher prediction accuracy. Through MMNet's adaptive learning mechanism, the genetic architecture of unknown traits can be analyzed.

[0009] 3. Capturing nonlinear effects of MMNet:

[0010] One of the highlights of MMNet is its ability to capture complex nonlinear effects. Specifically, in the prediction of plant height and heading date, it can capture information beyond additive heritability.

[0011] This application develops a new deep learning model based on genotype prediction of phenotype, and proposes a nonlinear deep learning network (MMNet) similar to a hybrid model, which systematically solves the shortcomings of existing technical solutions based on genotype prediction of phenotype.

[0012] 1. Existing mainstream methods, such as BLUP and GBLUP, usually rely on linear models. Although they have achieved certain results in the prediction of certain phenotypes, these methods have failed to effectively capture the nonlinear effects in complex traits. Although deep learning methods have been introduced into the field of genomic selection in recent years to improve this, existing deep learning models generally have the following problems: genetic background information and genomic variation information are not fully integrated. Current deep learning models usually treat the effects of genetic background information and genomic variation as independent input variables, and fail to effectively incorporate the complex interaction effects between the two into the prediction model. This processing method results in the model being unable to fully utilize the intrinsic relationship between genetic variation and background information, thus affecting the prediction accuracy.

[0013] We evaluated the performance of the MMNet model in predicting 12 agronomic traits of rice (such as yield, protein content, and heading date). The results showed that MMNet outperformed the existing six models (XGBoost, MLP, RF, GBLUP_Gs, BayesC, and DNNGP) in all traits. In terms of prediction performance, MMNet not only achieved the highest R 2 The scores of MMNet and its average and median Pearson correlation coefficients (PCC) far exceeded those of other models. MMNet also performed well in terms of error indicators (such as NMSE, RRMSE, and MAPE), especially when predicting complex traits such as yield, protein content, and heading date, with the smallest error. Overall, MMNet demonstrated strong adaptability and accuracy, and was able to effectively cope with the diversity and complexity of rice agronomic traits.

[0014] 2. Lack of adaptability: Existing methods lack the ability to adaptively adjust the contribution of genetic background information and genomic variation. Because different traits often have completely different genetic architectures. Some traits, such as the growth period of rice grown in a single production area and many rare genetic diseases in humans, are often controlled by a small number of major effect genes. There are also many traits, such as rice yield or human height, type 2 diabetes, etc., which are often controlled by a large number of small effect genes. For traits with different genetic architectures, the contribution ratio of genetic background and genomic variation to prediction accuracy is different. The existing model fails to dynamically adjust the contribution of the two according to the characteristics of different phenotypes, so that the prediction effect cannot be optimized. Further, the existing methods ignore the potential importance of the genetic structure of the phenotype and the genetic background information, and lack the ability to analyze the genetic structure of the traits.

[0015] We explored the contribution of ESN and VE subnetworks in MMNet to trait prediction. The results showed that the genetic structure of the trait determines the weight distribution of genetic background information and genomic variation information. For traits with similar genetic basis, such as heading date, plant height and stem length, the prediction mainly depends on key genomic variation, and the weight of VE is greater than that of ESN. For traits affected by a large number of small-effect genes, such as grain yield and protein content, MMNet relies more on genetic background information and improves prediction performance through the cumulative effects of small-effect genes. This weight distribution reflects the genetic structure of the trait and has important significance for genetics and breeding practice. The adaptive learning of MMNet is highly consistent with breeding practice and can provide valuable guidance for breeding of complex traits.

[0016] In general, this application proposal aims to solve these problems in existing methods and proposes a nonlinear deep learning network (MMNet) similar to a mixed model. During the MMNet training process, the contribution of genetic background information and genomic variation is adaptively adjusted, effectively analyzing the genetic structure of each trait and improving the accuracy of whole genome selection. Summary of the invention

[0017] The purpose of the present invention is to provide a model method for predicting biological traits by integrating genetic background and variation information, and proposes a nonlinear deep learning network (MMNet) quasi-hybrid model. During the MMNet training process, the contribution of genetic background information and genomic variation is adaptively adjusted, the genetic structure of each trait is effectively analyzed, and the accuracy of whole genome selection is improved.

[0018] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0019] A model for predicting biological traits by integrating genetic background and variation information, the model is a nonlinear deep learning network model similar to a mixed model, and includes the following modules:

[0020] Embedded Siamese network, which generates and extracts the genetic background and genetic correlation information between samples, and establishes an embedding vector for each sample to represent its genetic background information;

[0021] Genetic variation embedding network, which uses deep learning network to extract complex high-order information from the whole genome variation information of each sample and build embedding vectors to characterize the cumulative effect of its genetic variation on traits;

[0022] The fusion and trait prediction module achieves accurate prediction of traits by organically integrating the embedding vectors output by the above two modules.

[0023] Furthermore, the embedded Siamese network includes a backbone subnetwork and a weight sharing mechanism; the backbone subnetwork is composed of multiple fully connected layers, batch normalization layers, Leaky ReLU activation layers, and Dropout layers, and the input is genotype and phenotype data, and the network is trained using label values; the weight sharing mechanism generates an embedding vector for each sample based on the backbone subnetwork through weight sharing; the genetic correlation between samples is measured by the Euclidean distance between the embedding vectors, and the genetic correlation between a sample and the remaining samples is further generated by a multi-layer perceptron to generate corresponding embedding vectors.

[0024] Furthermore, the specific calculation formula for the genetic correlation between samples is as follows:

[0025]

[0026] Among them, qi and qj represent the embedding vectors of sample i and sample j respectively, and dis represents the Euclidean distance between the embedding vectors.

[0027] Furthermore, in the embedded Siamese network module, the genetic background information between samples provided by the user is directly used to generate relevant embedding vectors.

[0028] Furthermore, the genetic variation embedding network consists of two branches. The upper layer uses a residual unit, which consists of two 1×3 convolutional layers. A Batch Normalization layer and an activation function are inserted after the first convolution, and a final Batch Normalization layer is inserted after the second convolution; the jump connection consists of a 1×1 convolutional layer and a batch normalization layer; the lower branch includes a fully connected layer, a batch normalization layer, a Leaky ReLU activation layer and a Dropout layer; and finally, the information of the upper and lower layers is fused by weighted summation.

[0029] Furthermore, two scales of embedding vectors are generated in the embedded Siamese network and the genetic variation embedded network. The fusion and trait prediction module fuses the embedded representations of the embedded Siamese network and the genetic variation embedded network, and generates trait prediction results through multiple fully connected layers.

[0030] A method for predicting biological traits by integrating genetic background and variation information, comprising the following steps:

[0031] Step 1: Data preprocessing:

[0032] The preprocessed genetic data and corresponding phenotypic data are used for subsequent model training;

[0033] Step 2: Build a network model as in the above solution:

[0034] The network model is divided into three parts: embedded Siamese network, genetic variation embedded network, fusion and trait prediction module;

[0035] Step 3: Train the model and evaluate the contribution of the model training process.

[0036] Based on the principal component analysis algorithm or network weights, the contributions of the two sub-networks, the embedded Siamese network and the genetic variation embedded network, are evaluated.

[0037] Furthermore, the principal component analysis algorithm is as follows: for N samples in the data set, all embedding vectors Z1,…,ZN form a matrix Z of size N×2p; principal component analysis PCA based on singular value decomposition SVD is applied to the matrix Z; the contribution of each principal component to the original feature is calculated, and the weights of ESN and VE are weighted summed based on the variance contribution, so as to obtain the contribution weight W of each sub-network to the final prediction result VE and W ESN :

[0038]

[0039] U and V are the matrices obtained from the full SVD of the matrix Z, where U contains the left singular vectors and V contains the right singular vectors, is the i-th eigenvector V i The jth component of i is the i-th eigenvalue.

[0040] A computer device comprises a memory, a processor and computer-readable instructions stored in the memory and executable on the processor, wherein the method in the above scheme is implemented when the processor executes the computer-readable instructions.

[0041] A readable storage medium storing computer-readable instructions, wherein when the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps in the above scheme.

[0042] The beneficial effects of the present invention are as follows:

[0043] (1) Improved prediction accuracy: By integrating genetic background and gene variation information, MMNet can better capture nonlinear effects and complex genetic structures, thereby improving the prediction ability of complex traits.

[0044] (2) Shorten the breeding cycle: MMNet’s efficient prediction capability makes breeding decisions more accurate, thereby greatly shortening the breeding period.

[0045] (3) Analyzing the genetic architecture of phenotypes: MMNet can not only predict phenotypes, but also reveal the genetic structure of phenotypes, helping to deeply understand the complex relationship between genotype and phenotype. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a detailed introduction to the MMNet network. Figure a is an overview of the MMNet network; an extended Siamese network is used to extract the embedding of genetic background information; at the same time, a CNN-based genetic variation embedding network is used to extract the embedding of genomic variation information; the embeddings of the two layers are aligned and integrated for network training of the fusion layer, and finally the prediction of traits is achieved. Figure b is a detailed introduction to the ESN module in Figure a; a simple neural network is trained using genotype data and phenotypic data, and the embedded representations of samples in the network are extracted one by one. The genetic relationship between samples is obtained by evaluating the embeddings of different samples, and then the embedding vector representing the genetic relationship is obtained. Figure c is a detailed introduction to the residual unit used in the VE module in Figure a; it consists of two convolutions combined with batch normalization and Leaky ReLU activation functions, and the jump connection part directly processes the input data by a convolution followed by a batch normalization.

[0047] Figure 2 Compared with other commonly used methods, MMNet performs well in multiple rice agronomic traits. The bar chart in Figure a shows the R values ​​of 12 rice phenotypes predicted by various models. 2 The maximum value indicates that it has the highest prediction accuracy; the box plot in Figure b shows that MMNet has the highest mean and median of the correlation coefficients among the 12 phenotypes, indicating the robustness of its model; Figure c shows that the phenotypic errors are the smallest under the three different genetic architectures.

[0048] Figure 3 The predictive ability of plant height and growth period exceeded the additive heritability of the corresponding phenotype, indicating that nonlinearity was detected.

[0049] Figure 4 It is the dynamic pattern of the contribution of ESN embedding and VE embedding during training. Among the 12 agronomic traits, the VE (genomic variation information) contribution of heading date, plant height and stem length is significantly higher than the ESN (genetic background information) contribution. The ESN contribution of other traits is much higher than the VE contribution, which matches the phenotypic genetic architecture.

[0050] Figure 5 The prediction ability and PVE of MMNet for 12 agronomic traits change with LD threshold. a. The changing trend of R2 (red line) and PVE (blue line) when the LD threshold is in the range of 0.1 to 0.5. b. The correlation coefficient bar graph of the prediction ability (R2) and PVE of 12 traits. c. The prediction ability (R2) and PVE growth percentage bar graph of three LD threshold intervals (0.1→0.5, 0.1→0.3, 0.3→0.5). It shows that the prediction ability of the model is regulated by the genotype's explained variance of the phenotype. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0052] Embodiment 1

[0053] A model for predicting biological traits by integrating genetic background and variation information, specifically a nonlinear deep learning network (MMNet) model similar to a mixed model is proposed, which realizes the accurate prediction of animal and plant phenotypes and human physiological indicators or disease conditions using whole genome genetic variation; the model includes the following modules:

[0054] Embedding Siamese Network (ESN), which generates and extracts the genetic background and genetic correlation information between samples, and establishes an embedding vector for each sample to represent its genetic background information;

[0055] The Variation Embedding Network (VE) uses a deep learning network to extract complex high-order information from the whole genome variation information of each sample and build an embedding vector to characterize the cumulative effect of its genetic variation on the trait;

[0056] The fusion and trait prediction module achieves accurate prediction of traits by organically integrating the embedding vectors output by the above two modules.

[0057] The embedding Siamese network consists of a backbone subnetwork and a weight sharing mechanism. The backbone subnetwork consists of multiple fully connected layers, batch normalization layers, leaky ReLU activation layers, and dropout layers. The input is genotype (1×M) and phenotype data, and the network is trained using label values; while the weight sharing mechanism generates an embedding vector (1×T) for each sample based on the backbone subnetwork through weight sharing. The genetic correlation (GR) between samples is measured by the Euclidean distance between the embedding vectors. The genetic correlation between a sample and the rest of the samples is further generated through a multi-layer perceptron to generate the corresponding embedding vector.

[0058] In the Embedded Siamese Network module, the genetic background information between samples provided by the user can be directly used to generate relevant embedding vectors.

[0059] The genetic variation embedding network consists of two branches. The upper layer uses a residual unit, which consists of two 1×3 convolutional layers. The first convolution layer is followed by a Batch Normalization layer and an activation function, and the second convolution layer is followed by a final Batch Normalization layer. The Skip Connection consists of a 1×1 convolutional layer and a Batch Normalization layer. The lower branch includes a fully connected layer, a Batch Normalization layer, a Leaky ReLU activation layer, and a Dropout layer. Finally, the information of the upper and lower layers is fused by weighted summation.

[0060] Embodiment 2

[0061] Step 1: Data preprocessing;

[0062] First, the rice genotype and phenotype data were quality controlled (QC) to ensure the integrity of the data and the reliability of the analysis. Data quality control includes the following three steps: Variant screening: First, genetic markers with a missing value rate of more than 25% were screened out; then genetic markers with a MAF (minor allele frequency) lower than 0.05 were excluded. The genetic variation data after quality control were encoded with 1 and -1 for homozygous and heterozygous genotypes, respectively, and missing genotype values ​​were filled with 0.

[0063] We conducted a normal distribution test on the phenotypic data, and used specific mathematical transformations to transform the relevant phenotypic data that did not meet the normal distribution so that they met the normal distribution. The 12 rice agronomic traits presented in this application all passed the normal distribution test.

[0064] The preprocessed genetic data and corresponding phenotypic data are used for subsequent model training.

[0065] Step 2: Build the MMnet network;

[0066] The MMnet network is mainly divided into three parts: Embedding Siamese Network (ESN), Variation Embedding Network (VE), and Embedding Fusion Module. The following is a detailed introduction.

[0067] (1) ESN module

[0068] This embodiment proposes a new Siamese network architecture, which enhances the network's ability to learn the gene correlation between samples by introducing embedding and weight sharing mechanisms. The ESN module includes two main parts.

[0069] Backbone subnetwork: The backbone subnetwork is used to embed the training samples, and the input includes the genotype and phenotype data of each sample. This subnetwork consists of multiple fully connected layers, batch normalization layers, leaky ReLU activation layers, and dropout layers. The input is genotype (1×M) and phenotype data, and the network is trained using label values.

[0070] Weight sharing mechanism: Based on the backbone sub-network, an embedding vector (1×T) is generated for each sample through a weight sharing mechanism. The genetic relevance (GR) between samples is measured by the Euclidean distance between the embedding vectors. The specific calculation formula is as follows:

[0071]

[0072] Among them, q i and q j They represent the embedding vectors of sample i and sample j respectively, and dis represents the Euclidean distance between the embedding vectors.

[0073] (2) VE module

[0074] The VE module processes genotype data and extracts genomic information through two branches.

[0075] Upper branch: The upper layer uses a residual unit, which consists of two 1×3 convolutional layers. The first convolution is followed by a Batch Normalization layer and an activation function, and the second convolution is followed by a final BatchNormalization layer. Skip Connection consists of a 1×1 convolutional layer and a batch normalization layer. The convolution operation of the upper layer effectively captures the local features in the genotype data and reduces the gradient vanishing problem through the residual unit.

[0076] Lower layer branch: The lower layer includes a fully connected layer, a batch normalization layer, a Leaky ReLU activation layer, and a Dropout layer. Finally, the information of the upper and lower layers is fused by weighted summation to output a 1024-dimensional embedding based on gene mutation information.

[0077] The function of the VE module is to convert the input genotype data from (N×M) to an embedded representation of (N×1024), extract complex gene variation information, and enhance the information mining capability of the network.

[0078] (3) Embedded Fusion Module

[0079] The embedding fusion module concatenates the embedding vectors (1×p) from the ESN and VE sub-networks to form a new vector (1×2p). This fused vector is batch normalized and then input into the scoring unit for final prediction. The scoring unit includes multiple fully connected layers, Leaky ReLU activation layers, and Dropout layers, and finally outputs the predicted phenotypic value.

[0080] During the training process, in order to ensure that the contributions of genetic background information and genomic variation information are fairly evaluated, in addition to calculating the loss through the final prediction value, the embedding of each sub-network is passed through the scoring unit and added to the final loss, thereby training the network more comprehensively.

[0081] Step 3: Contribution evaluation during model training;

[0082] In order to evaluate the contribution of the two sub-networks ESN and VE, this embodiment proposes an algorithm based on principal component analysis (PCA). Specifically, for N samples in the dataset, all embedding vectors Z 1 ,…,Z N A matrix Z of size (N×2p) is formed. Principal component analysis (PCA) based on singular value decomposition (SVD) is applied to the matrix Z. The contribution of each principal component to the original features is calculated, and the weights of ESN and VE are weighted summed based on the variance contribution to obtain the contribution weight (W) of each sub-network to the final prediction result. VE and W ESN ).

[0083]

[0084] U and V are the matrices obtained from the full SVD of the matrix Z, where U contains the left singular vectors and V contains the right singular vectors, is the jth component of the i-th eigenvector Vi, and ηi is the i-th eigenvalue.

[0085] The solution of this patent is based on the intersection of deep learning and genetics, and improves the accuracy of phenotypic prediction by combining nonlinear modeling methods (such as Siamese networks, convolutional neural networks, residual connections, etc.) with statistical analysis methods such as PCA. This solution can adapt to different genomic data and phenotypic characteristics, dynamically adjust the contribution of genetic background and variation information, and thus show excellent performance in the prediction of a variety of complex traits.

[0086] Through the above technical solutions, the present invention can solve the problem in the prior art that genetic background information and genomic variation information cannot be effectively integrated, and effectively capture the nonlinear effects between genes, greatly improving the accuracy of genomic selection and the ability to analyze the genetic architecture of complex traits.

[0087] In summary, the present invention can be applied to the following aspects:

[0088] Animal and plant breeding: It can be widely used in whole genome selection to achieve efficient breeding of crops and animals.

[0089] Genetic Research: MMNet provides geneticists with a new method to analyze the genetic architecture of traits and can fully utilize nonlinear effects to assist trait prediction.

[0090] Biological trait prediction: MMNet can be used to predict phenotypes of different plants or animals based on genomic information, especially complex traits (quality, yield, etc.).

[0091] Disease prediction: MMNet can predict the onset patterns of specific diseases and physiological indicators and other traits based on human genome information.

[0092] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the protection scope of the present invention in any form, and all technical solutions obtained by equivalent replacement and the like fall within the protection scope of the present invention. The parts not involved in the present invention are the same as the prior art or can be implemented by the prior art.

Claims

1. A model for predicting biological traits by integrating genetic background and variation information, characterized in that: The model is a nonlinear deep learning network model similar to a hybrid model, including the following modules: Embedded Siamese network, which generates and extracts the genetic background and genetic correlation information between samples, and establishes an embedding vector for each sample to represent its genetic background information; Genetic variation embedding network, which uses deep learning network to extract complex high-order information from the whole genome variation information of each sample and build embedding vectors to characterize the cumulative effect of its genetic variation on traits; The fusion and trait prediction module achieves accurate prediction of traits by organically integrating the embedding vectors output by the above two modules.

2. A model for predicting biological traits by integrating genetic background and variation information according to claim 1, characterized in that: The embedded Siamese network includes a backbone subnetwork and a weight sharing mechanism; the backbone subnetwork is composed of multiple fully connected layers, batch normalization layers, leaky ReLU activation layers, and dropout layers, and the input is genotype and phenotype data, and the network is trained using label values; the weight sharing mechanism generates an embedding vector for each sample based on the backbone subnetwork through weight sharing; the genetic correlation between samples is measured by the Euclidean distance between the embedding vectors, and the genetic correlation between a sample and the remaining samples is further generated by a multi-layer perceptron to generate a corresponding embedding vector.

3. A model for predicting biological traits by integrating genetic background and variation information according to claim 2, characterized in that: The specific calculation formula for the genetic correlation between samples is as follows: Among them, qi and qj represent the embedding vectors of sample i and sample j respectively, and dis represents the Euclidean distance between the embedding vectors.

4. A model for predicting biological traits by integrating genetic background and variation information according to claim 1, characterized in that: In the Embedded Siamese Network module, the genetic background information between samples provided by the user is directly used to generate relevant embedding vectors.

5. The model for predicting biological traits by integrating genetic background and variation information according to claim 1, characterized in that: The genetic variation embedding network consists of two branches. The upper layer uses a residual unit, which consists of two 1×3 convolutional layers. The Batch Normalization layer and activation function are inserted after the first convolution, and the second convolution is followed by a final Batch Normalization layer; the jump connection consists of a 1×1 convolutional layer and a batch normalization layer; the lower branch includes a fully connected layer, a batch normalization layer, a Leaky ReLU activation layer and a Dropout layer; finally, the information of the upper and lower layers is fused by weighted summation.

6. A model for predicting biological traits by integrating genetic background and variation information according to claim 1, characterized in that: In the embedded Siamese network and the genetic variation embedded network, two-scale embedding vectors are generated, and the trait prediction results are generated by embedding fusion.

7. A method for predicting biological traits by integrating genetic background and variation information, characterized in that: The following steps are involved: Step 1: Data preprocessing: The preprocessed genetic data and corresponding phenotypic data are used for subsequent model training; Step 2: construct a network model as described in any one of claims 1 to 6: The network model is divided into three parts: embedded Siamese network, genetic variation embedded network, fusion and trait prediction module; Step 3: Train the model and evaluate the contribution of the model training process. Based on the principal component analysis algorithm or network weights, the contributions of the two sub-networks, the embedded Siamese network and the genetic variation embedded network, are evaluated.

8. The method for predicting biological traits by integrating genetic background and variation information according to claim 7, characterized in that: The principal component analysis algorithm is as follows: for N samples in the data set, all embedding vectors Z1,…,ZN form a matrix Z of size N×2p; principal component analysis PCA based on singular value decomposition SVD is applied to the matrix Z; the contribution of each principal component to the original feature is calculated, and the weights of ESN and VE are weighted summed based on the variance contribution, so as to obtain the contribution weight W of each sub-network to the final prediction result VE and W ESN : U and V are the matrices obtained from the full SVD of the matrix Z, where U contains the left singular vectors and V contains the right singular vectors, is the i-th eigenvector V i The jth component of i is the i-th eigenvalue.

9. A computer device comprising a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, characterized in that: When the processor executes the computer-readable instructions, the method according to any one of claims 7 to 8 is implemented.

10. A readable storage medium storing computer-readable instructions, wherein when the computer-readable instructions are executed by one or more processors, the one or more processors are caused to perform the steps of any one of claims 7 to 8.

Images