Breeding phenotype prediction method and system for genetic sequence data

Abstract

The present application relates to the technical field of breeding phenotype prediction, and discloses a breeding phenotype prediction method and system for gene sequence data, which comprises the following steps: acquiring genotype data and phenotype data to construct a data set; dividing the data set into a training set and a test set; constructing a convolution operator space, a pooling operator space and a neural network node, assigning convolution operators and pooling operators to the neural network node, and generating a neural network structure; converting the neural network structure into a tree data structure and optimizing it by using a depth-first search algorithm to generate an executable neural network model; training the executable neural network model by using the training set, and optimizing the trained neural network model by using a meme algorithm to obtain an optimal breeding phenotype prediction model; and inputting test set data into the optimal breeding phenotype prediction model for processing to obtain a breeding phenotype prediction result. The present application significantly improves the prediction accuracy in the breeding phenotype prediction of complex traits.

Description

Technical Field

[0001] This invention relates to the field of breeding phenotype prediction technology, and more specifically, to a breeding phenotype prediction method and system based on gene sequence data. Background Technology

[0002] With the continuous growth of the global population and the constant changes in environmental conditions, food security is becoming an increasingly serious issue. Modern breeding technologies, by improving the genetic characteristics of species and enhancing their yield, disease resistance, nutritional quality, and environmental adaptability, provide an important pathway to solving food shortages. The rapid development of precision breeding and gene editing technologies has made it possible to mine genetic diversity and introduce desirable traits at the genomic level. However, the genetic improvement of complex traits (such as feed conversion ratio and stress resistance) faces many challenges, and there is an urgent need to develop efficient prediction and selection methods to shorten the breeding cycle and improve breeding efficiency.

[0003] Current technologies primarily employ genomic selection (GS) methods, which predict an individual's genetic potential by combining genotype and phenotype information. In recent years, deep learning-based genomic selection methods, such as Deep Learning Genomic Selection (DeepGS) and dual-CNN stream genome-wide association studies, have gradually become research hotspots. These methods extract key features from high-throughput sequencing data and deep neural network models, enabling them to capture the complex relationships between genotype and phenotype, thus improving the efficiency and accuracy of complex trait prediction. Furthermore, some studies have combined convolutional neural networks (CNNs) with fully connected layers to handle multigene and environment interaction effects (GxE), further enhancing the predictive power for complex traits.

[0004] However, traditional GS methods struggle to capture the nonlinear relationship between genotype and traits and fail to adequately consider gene-environment interactions, resulting in low prediction accuracy for complex traits. Meanwhile, deep learning-based models (such as the DNNGP method) also face numerous challenges in practical applications: they lack the ability to process multi-channel data input, are insufficiently robust to input data shapes, have limited high-level feature extraction capabilities, and their overly simplistic network structures make them ill-suited for complex breeding needs. These shortcomings directly limit the prediction accuracy for complex traits. Summary of the Invention

[0005] To improve the prediction accuracy of breeding phenotypic prediction technology, this invention proposes the following technical solution: In a first aspect, this invention proposes a breeding phenotype prediction method based on gene sequence data, comprising: Acquire genotype and phenotypic data, and construct a dataset using the genotype and phenotypic data; The dataset is divided into a training set and a test set; Construct the convolution operator space, pooling operator space, and neural network nodes; From the convolution operator space and the pooling operator space, convolution operators and pooling operators are assigned to neural network nodes to generate a neural network structure; The neural network structure is converted into a tree data structure, and the tree data structure is optimized using a depth-first search algorithm to generate an executable neural network model. The executable neural network model is trained using the training set, and the trained neural network model is optimized using the meme algorithm to obtain the optimal breeding phenotype prediction model. The test set data is input into the optimal breeding phenotype prediction model for processing to obtain the breeding phenotype prediction results.

[0006] As a preferred technical solution, after dividing the dataset into a training set and a test set, the method further includes: For even-dimensional feature data in the training set, the encoding is performed according to the following formula:

[0007] For odd-dimensional feature data in the training set, the encoding is performed according to the following formula:

[0008] in, Indicates the location index of the sample data. The dimension index of the feature vector. Indicates the dimension of the feature vector.

[0009] As a preferred technical solution, the convolution operator space and pooling operator space are constructed according to the following formula:

[0010] in, It represents a set of characters, where each character is used to identify a convolution operator or a pooling operator; For the set of convolution operators, characters Different convolution operators, characters This indicates no connection or no operation. For the pooling operator set, characters Different pooling operators, To store the set of operators selected from the convolution operator space or the pooling operator space, For the convolution operator space, For the pooling operator space, This indicates that characters are mapped to the corresponding convolution operators or pooling operations.

[0011] As a preferred technical solution, the neural network node includes a CBAN attention mechanism and multiple convolutional blocks with residual connections; The CBAN attention mechanism includes channel attention mechanism and spatial attention mechanism; The channel attention mechanism is used to weight the different channels of the input data, including: The channel attention weights are calculated according to the following formula. :

[0012] in, The feature map representing the input data, For the number of channels, and The height and width of the feature map; and For different learning weight matrices, For the reduction rate, For activation function, Sigmoid activation function; Generate the channel attention weight feature map according to the following formula. :

[0013] in, This represents element-wise dot product; The spatial attention mechanism is used to weight the spatial location of the input data, including: The spatial attention weights are calculated according to the following formula. :

[0014] in, Indicates the feature map Average pooling is performed along the channel dimension. This represents max pooling of the feature maps along the channel dimension, while concat() represents concatenating two feature maps along the channel dimension. Generate a spatial attention map according to the following formula. :

[0015] The output of each neural network node is generated according to the following formula:

[0016] in, Indicates the first The nth neural network node is related to the original input or the nth neural network node. The connection method of each neural network node This refers to the index position information of the neural network node, used to determine the first... Each neural network node and Is there a connection between them? This represents the computational operation unit of the current neural network node.

[0017] As a preferred technical solution, the neural network structure is converted into a tree data structure, and the tree data structure is optimized using a depth-first search algorithm to generate an executable neural network model, including: By traversing the layers of the neural network structure, the nodes of the neural network structure are defined as nodes in a tree data structure, and each node is represented as a triple. The root node is defined as a node with no input operations, and the leaf nodes are defined as terminal nodes in the tree data structure with no child nodes. The value of each leaf node contains the set of operators and neural network nodes corresponding to that leaf node, thus constructing the initial tree data structure. The value of the node represents the state of the neuron; For the index of the node; It is a set of child nodes; Starting from the root node, recursively traverse all child nodes, recording the set of all paths from the root node to a leaf node. :

[0018] in, For path i The node values ​​in The length of the path; For path set Add dimensionality reduction layers and linear layers to obtain an executable neural network model.

[0019] As a preferred technical solution, a loss function is constructed, and the executable neural network model is trained using an optimizer and a learning rate scheduler, including: Using the smooth L1 loss function as the training loss function, its expression is as follows:

[0020] In the formula, , For the true value, These are the model's predicted values; The Adam optimizer is used to update the training parameters of the executable neural network model, including: Calculate the current value according to the following formula. t First moment estimation at time 1 , representing the exponentially weighted average of the gradient:

[0021] in, The exponentially weighted decay rate of the first moment. For the current gradient, for t- First-order moment estimate at time 1; Calculate the current value according to the following formula. t Second moment estimation at time 1 , representing the exponentially weighted average of the squared gradient:

[0022] in, The exponentially weighted decay rate of the second moment; The first moment is estimated according to the following formula. and estimation of second moment Perform deviation correction processing:

[0023]

[0024] in, This is the deviation correction value for the first moment. This is the deviation correction value for the second moment. for of t Power of 1 for of t Power of; The training parameters of the executable neural network model are updated according to the following formula:

[0025] in, For the first t The training parameter values ​​after the next iteration For the first The training parameter values ​​at the next iteration For learning rate, To prevent division by zero of constants; The cosine annealing learning rate scheduler is used to update the learning rate of an executable neural network model, as shown in the following expression:

[0026] in, Let be the learning rate for cosine annealing, and let represent the learning rate at the th... t The learning rate used during each training session. To minimize the learning rate, The initial learning rate, For training cycles, This is the current training round.

[0027] As a preferred technical solution, after training the executable neural network model and before optimizing the trained neural network model using a meme algorithm, the method further includes: The Pearson coefficient of the trained neural network model is calculated as the initial fitness value for the meme algorithm according to the following formula:

[0028] in, For the number of samples, the variable values These are the predicted values ​​of the neural network model, representing variable values. This represents the true breeding phenotypic label value. for and The sum of the products of corresponding values For variable values The sum, For variable values The sum, For variable values The sum of squares, For variable values The sum of squares.

[0029] As a preferred technical solution, a meme algorithm is used to optimize the trained neural network model to obtain the optimal breeding phenotype prediction model, including: Using two trained neural network models as individuals in the population, individuals are selected from the population using a roulette wheel selection algorithm based on their initial fitness values. Perform crossover operations on the selected individuals to generate new candidate individuals; Perform mutation operations on new candidate individuals; Perform a local search on the individuals after mutation and recalculate their fitness values; The individual with the highest fitness value is selected iteratively using the elite selection algorithm and the tournament selection algorithm. The neural network model of the individual with the highest fitness value is then used as the optimal breeding phenotype prediction model.

[0030] As a preferred technical solution, individuals are selected from the population using a roulette wheel selection algorithm based on initial fitness values, including: Calculate the sum of fitness values ​​for all individuals using the following formula. :

[0031] in, Indicates the first Individual, Indicates calculation fitness value; Calculate the individual according to the following formula. Probability of being selected :

[0032] Construct the cumulative probability distribution according to the following formula. :

[0033] In the formula, Indicates the selection of individuals The cumulative probability, Generate random numbers To obtain satisfaction individual As the selected individual; Perform crossover operations on the selected individuals to generate new candidate individuals, including: The two individuals selected will be designated as parent individuals. and parental individuals ; Select the intersection point and alternately swap the parent individuals. and parental individuals The gene segment is used to generate two offspring individuals as new candidate individuals; Mutation operations are performed on new candidate individuals, including: Randomly select a gene locus from a new candidate individual and change the value of that gene locus to another value, or randomly select two gene loci between two new candidate individuals and exchange them; Perform a local search on the individuals after mutation, and recalculate the fitness values ​​of the individuals, including: For each individual after mutation, a separation operation is performed on the neural network model to obtain a convolutional information stream containing only convolution operators and a pooling information stream containing only pooling operators; The pooling information stream is iteratively mutated, and the mutated pooling information stream and convolutional information stream are merged to obtain a new neural network model. New individuals are created using a novel neural network model, and their fitness values ​​are calculated.

[0034] Secondly, the present invention also proposes a breeding phenotype prediction system based on gene sequence data, applied to the breeding phenotype prediction method based on gene sequence data as described in any embodiment of the first aspect, comprising: The acquisition module is used to acquire genotype data and phenotypic data, and to construct a dataset using the genotype data and phenotypic data; The dataset is divided into training and testing sets by partitioning it into modules. The building blocks are used to construct the convolution operator space, pooling operator space, and neural network nodes; The allocation module is used to allocate convolution operators and pooling operators to neural network nodes from the convolution operator space and the pooling operator space to generate a neural network structure; The conversion and optimization module is used to convert the neural network structure into a tree data structure and optimize the tree data structure using a depth-first search algorithm to generate an executable neural network model. The training optimization module is used to train the executable neural network model using the training set, and to optimize the trained neural network model using the meme algorithm to obtain the optimal breeding phenotype prediction model. The prediction module is used to input the test set data into the optimal breeding phenotype prediction model for processing, and obtain the breeding phenotype prediction results.

[0035] The beneficial effects of the present invention include at least the following: This invention constructs convolution and pooling operator spaces to dynamically allocate suitable convolution and pooling operators to neural network nodes, generating an efficient and highly adaptable neural network structure. It also utilizes a depth-first search (DFS) algorithm to optimize the tree-like data structure of the neural network, ensuring the accuracy and efficiency of feature extraction paths. Combining meme algorithms with global and local optimization of the network model not only improves the model's convergence speed but also significantly enhances its predictive ability for complex traits (such as feed conversion ratio, yield, and disease resistance). Through optimized design, the model can extract more biologically meaningful nonlinear relationship features from high-dimensional genotype and phenotypic data, effectively capturing the complexity of multigene and environment interaction effects (GxE), providing strong support for the accurate prediction of complex traits.

[0036] Meanwhile, this invention dynamically adjusts the connection structure of neural network nodes to ensure consistency between input data and network structure, enhancing the model's robustness to changes in input data and avoiding prediction errors caused by data anomalies or shape mismatches. The model employs an advanced feature extraction mechanism combining multi-layer convolution and pooling to effectively capture key biological signals in the input data, significantly improving the extraction efficiency and accuracy of features required for complex trait prediction. Through optimization on the training set and validation on the test set, the model demonstrates good adaptability and scalability, enabling flexible application to different scenarios and datasets, ensuring the broad applicability and stability of complex trait prediction.

[0037] Ultimately, this invention significantly improves the prediction accuracy of breeding phenotypes for complex traits (such as feed conversion ratio, yield, and disease resistance), overcomes the technical bottlenecks of traditional genomic selection methods in capturing complex interaction effects and predicting low accuracy, and provides reliable technical support for modern breeding and genomics research. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the breeding phenotype prediction method based on gene sequence data provided in an embodiment of the present invention.

[0039] Figure 2 This is an example diagram illustrating the connection and stacking of operators and neural network nodes in an embodiment of the present invention.

[0040] Figure 3 This is a schematic diagram of the operator and neural network node tree construction process in an embodiment of the present invention.

[0041] Figure 4 This is a flowchart illustrating the execution of the depth-first search algorithm according to an embodiment of the present invention.

[0042] Figure 5 This is a schematic diagram illustrating the change of the loss function during the training process in an embodiment of the present invention.

[0043] Figure 6 This is a schematic diagram illustrating the process of executing the meme algorithm in an embodiment of the present invention.

[0044] Figure 7 This is a flowchart illustrating the breeding phenotype prediction system based on gene sequence data provided in an embodiment of the present invention. Detailed Implementation

[0045] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred technical solutions. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred technical solutions are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0046] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0047] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0048] Example 1 This embodiment proposes a breeding phenotype prediction method based on gene sequence data, such as... Figure 1 As shown, Figure 1 This is a flowchart illustrating a breeding phenotype prediction method based on gene sequence data provided in this embodiment. The method includes the following steps: S1: Obtain genotype data and phenotypic data, and construct a dataset using the genotype data and phenotypic data.

[0049] S2: Divide the dataset into training and test sets; S3: Construct the convolution operator space, pooling operator space, and neural network nodes; S4: Assign convolution operators and pooling operators to neural network nodes from the convolution operator space and the pooling operator space to generate a neural network structure; S5: Convert the neural network structure into a tree data structure, and optimize the tree data structure using a depth-first search algorithm to generate an executable neural network model; S6: Using the training set, train the executable neural network model, and use the meme algorithm to optimize the trained neural network model to obtain the optimal breeding phenotype prediction model. S7: Input the test set data into the optimal breeding phenotype prediction model for processing to obtain the breeding phenotype prediction results.

[0050] Understandably, this invention constructs convolution and pooling operator spaces to dynamically allocate suitable convolution and pooling operators to neural network nodes, generating an efficient and highly adaptable neural network structure. It also utilizes a depth-first search (DFS) algorithm to optimize the tree-like data structure of the neural network, ensuring the accuracy and efficiency of feature extraction paths. Combining meme algorithms with global and local optimization of the network model not only improves the model's convergence speed but also significantly enhances its predictive ability for complex traits (such as feed conversion ratio, yield, and disease resistance). Through optimized design, the model can extract more biologically meaningful nonlinear relationship features from high-dimensional genotype and phenotypic data, effectively capturing the complexity of multigene and environment interaction effects (GxE), providing strong support for the accurate prediction of complex traits.

[0051] Meanwhile, this invention dynamically adjusts the connection structure of neural network nodes to ensure consistency between input data and network structure, enhancing the model's robustness to changes in input data and avoiding prediction errors caused by data anomalies or shape mismatches. The model employs an advanced feature extraction mechanism combining multi-layer convolution and pooling to effectively capture key biological signals in the input data, significantly improving the extraction efficiency and accuracy of features required for complex trait prediction. Through optimization on the training set and validation on the test set, the model demonstrates good adaptability and scalability, enabling flexible application to different scenarios and datasets, ensuring the broad applicability and stability of complex trait prediction.

[0052] Ultimately, this invention significantly improves the prediction accuracy of breeding phenotypes for complex traits (such as feed conversion ratio, yield, and disease resistance), overcomes the technical bottlenecks of traditional genomic selection methods in capturing complex interaction effects and predicting low accuracy, and provides reliable technical support for modern breeding and genomics research.

[0053] Example 2 This embodiment is an improvement on the breeding phenotype prediction method based on gene sequence data proposed in Example 1.

[0054] In this embodiment, genotype and phenotypic data are obtained from public databases, such as high-throughput sequencing data and field phenotypic measurement data of crops like wheat and corn. During the data processing stage, phenotypic data (such as plant height, disease resistance, and feed conversion ratio) are standardized to ensure they fully reflect actual breeding objectives. The original data is reshaped into a shape of (2000, 1, 1691), where 2000 represents the number of samples, 1 represents the number of channels, and 1691 represents the feature dimension. The data is divided into training and test sets at a ratio of 0.9, with the training set having a shape of (1800, 1, 1691) and the test set having a shape of (200, 1, 1691). Position encoding is performed on the training set, fusing the sample position index with the feature vector dimension.

[0055] For even-dimensional feature data in the training set, the encoding is performed according to the following formula:

[0056] For odd-dimensional feature data in the training set, the encoding is performed according to the following formula:

[0057] in, Indicates the location index of the sample data. The dimension index of the feature vector. Indicates the dimension of the feature vector.

[0058] The encoded training set data is packaged using DataLoader. DataLoader has the following features: supports efficient loading of data from memory or disk, batching of data, avoiding bias caused by data order during training, and supporting multi-threading to accelerate data reading.

[0059] Understandably, by adding a unique code to each position in the sequence, the model can understand the order and positional information of the input data. This encoding not only helps the model capture relative positional information in the sequence but also improves the representational power of the sequence data. By combining sine and cosine in different dimensions, a unique high-dimensional vector is generated for each position. These vectors have good distribution and discriminability in high-dimensional space, helping the model distinguish inputs at different positions. Sine and cosine positional encoding does not rely on a fixed-length input sequence and can dynamically calculate positional codes of arbitrary length as needed, making it suitable for sequence data of various sizes. This flexibility significantly improves the model's versatility.

[0060] In this embodiment, the convolution operator space and the pooling operator space are constructed according to the following formula:

[0061] in, It represents a set of characters, where each character is used to identify a convolution operator or a pooling operator; For the set of convolution operators, characters Different convolution operators, characters This indicates no connection or no operation. For the pooling operator set, characters Different pooling operators, To store the set of operators selected from the convolution operator space or the pooling operator space, For the convolution operator space, For the pooling operator space, This indicates that characters are mapped to the corresponding convolution operators or pooling operations.

[0062] It's important to note that convolution operators can identify key gene segments in genotypic data that are related to complex traits. For example, when predicting wheat disease resistance, convolution operators can extract potential patterns in disease resistance gene locus regions. Pooling operators, by reducing the dimensionality of feature maps and compressing redundant information, enable models to maintain efficient computation when processing large-scale genotypic data and capture the macroscopic relationship between phenotypes and genotypes. Each convolution operator is encapsulated as a class, implementing convolution operations on the input data within the class. This encapsulation facilitates modular calling and flexible adjustment. Pooling operators do not require encapsulation into classes and can be used directly, reducing the complexity of the operator space. Key-value pairs map characters to operator objects in both the convolution and pooling operator spaces. The `random` method is then used to randomly return characters and corresponding operators from the convolution and pooling operator spaces to generate the model structure. Neural network nodes are predefined as classes, created cyclically according to the number of nodes, while the convolution kernel size is fixed. Figure 2 As shown, Figure 2 This is an example diagram illustrating the connection and stacking of operators and neural network nodes in an embodiment of the present invention. x The initial data for the model is passed and processed sequentially through multiple neural network nodes and operators. Node numbers (e.g., 0, 1, 2, 3) represent neural network nodes. Operators (e.g., a, b, c, d, e) connect different neural network nodes and are used to perform operations such as convolution or pooling. The arrows indicate the data flow paths. Through these paths, data is progressively passed from the input nodes to the output nodes.

[0063] In this embodiment, the neural network node includes a CBAN attention mechanism and multiple convolutional blocks with residual connections; The CBAN attention mechanism includes channel attention mechanism and spatial attention mechanism; The channel attention mechanism is used to weight the different channels of the input data, including: The channel attention weights are calculated according to the following formula. :

[0064] in, Feature maps representing input data; Number of channels; and The height and width of the feature map; and For different learning weight matrices; The reduction rate is 16 or 8. Use ReLU as the activation function; The Sigmoid activation function is used to compress the output to the range [0,1].

[0065] Generate the channel attention weight feature map according to the following formula. :

[0066] in, This represents element-wise dot product; The spatial attention mechanism is used to weight the spatial location of the input data, including: The spatial attention weights are calculated according to the following formula. This indicates the importance of each spatial location:

[0067] in, Indicates the feature map Average pooling along the channel dimension yields the following result: Feature map; This indicates that max pooling of the feature map along the channel dimension yields the same result. The feature maps; concat() means concatenating two feature maps along the channel dimension.

[0068] Generate a spatial attention map according to the following formula. : .

[0069] Understandably, CBAM (Channel and Spatial Attention Module) significantly improves task performance by adaptively adjusting channel and spatial weights, enabling the model to automatically focus on key feature regions. The CBAM module has a compact and simple design, allowing for easy integration into existing convolutional neural networks without significant modifications to the network architecture. This characteristic makes CBAM an efficient and easy-to-use enhancement module. By introducing the attention mechanism, CBAM can visually demonstrate the channels and spatial regions the network focuses on when processing input data, aiding in understanding the model's decision-making process and feature extraction procedures.

[0070] A neural network node consists of multiple convolutional blocks and uses residual connections. The input is added to the output after convolution by the CBAM attention mechanism and the neural network node to obtain the final output.

[0071] For the There are a total of neural network nodes, which have a total of Each information flow location can be used for convolution and pooling operators to connect. Among them, Represented as the current neural network node The connection method between the original input and the neural network nodes; if 0 represents the original input, then the... The output of each neural network node is as follows:

[0072] in, Indicates the first The nth neural network node is related to the original input or the nth neural network node. The connection method of each neural network node This refers to the index position information of the neural network node, used to determine the first... Each neural network node and Is there a connection between them? This represents the computational operation unit of the current neural network node.

[0073] Residual connections enhance gradient flow by skipping certain layers, mitigating the vanishing gradient problem during training. The definition of a residual connection is as follows:

[0074] in, This represents the input feature map; it is usually the output of the previous layer. This represents a learnable function, typically composed of multiple neural network layers, such as convolutional layers, activation functions (e.g., ReLU), and normalization layers (e.g., Batch Normalization). This represents the output feature map after residual connection, which includes the sum of the input feature map and the processed feature map.

[0075] Understandably, in deep networks, gradients can vanish during backpropagation, leading to training difficulties. Residual connections significantly alleviate this problem by allowing gradients to propagate directly from the output layer to the input layer, thus improving the feasibility of training deep networks. By learning residuals (i.e., the difference between the input and the target) rather than the complete mapping, residual connections reduce optimization difficulty, allowing deeper networks to be trained more efficiently. By simplifying the objective function, residual connections make it easier for the model to capture complex feature relationships, thereby improving overall performance. Residual connections allow features extracted by previous layers to be directly reused by subsequent layers, ensuring that important information is not lost. This feature reuse mechanism further enhances the model's expressive power.

[0076] In this embodiment, after generating the neural network structure, each neural network structure is decoded, and each operator in the individual is checked. If the operator does not match the data shape, the operator is reassigned. Specifically, this includes: Create an empty two-dimensional array of shape (Neural_nodes, Neural_nodes + 1), where each row represents a neural network node and each column represents the input information of the neural network node, including the connection information between neural network nodes and the selection state of operators. Iterate through the neural network structure to perform decoding operations. The final decoding result is a set of characters recording the information flow connections within each neural network structure (such as character identifiers for convolution or pooling operators).

[0077] Then, after decoding the initial model, an identifier matrix is ​​generated to label the operator type of each connection in the network structure, ensuring the model's execution and connection validity. Each row of the identifier matrix corresponds to a node, and each column is defined as follows: -1 represents a pooling operation, 1 represents a convolution operation, and 0 represents no connection. Based on the identifier matrix, connection information for each node is generated, and it is determined which nodes have convolutional or pooling connections. This ensures that it is possible to identify which connections are valid, which are unconnected, and which are the results of convolutional or pooling operations. Simultaneously, by combining the decoding results and the identifier matrix, each neural network structure set can be returned, clarifying the relationships between nodes and the overall network topology of the model.

[0078] A neural network layer is dynamically constructed from the neural network structure and the label matrix, avoiding frequent memory allocation. An outer loop controls the number of network construction iterations, iterating through each label matrix and storing the column information of each row in a `ModuleList`. The processing method for each position is determined based on the value (0, -1, 1, 2) in the label matrix. 0 indicates no layer; add a NoPlayer instance.

[0079] -1 indicates a pooling operation, which retrieves and adds layers from an existing network layer.

[0080] 1 and 2 represent convolution operations and neural network node operations. If the number of channels in the current layer does not meet the requirements, a new layer is created to ensure that the model adapts to the data flow.

[0081] Ultimately, a network structure consisting of multiple layers of neural networks was constructed, with the configuration of each layer flexibly adjusted according to the values ​​in the matrix, ensuring the scalability and performance of the model.

[0082] Understandably, convolutional operators focus on feature extraction, learning local features and patterns (such as edges and textures), while pooling operators reduce the spatial dimensionality of feature maps, decreasing computational cost while preserving important information. Separation allows for flexible adjustment of the parameters of convolutional and pooling layers, such as kernel size, stride, and pooling method, to adapt to various task requirements. Convolutional layers capture fine-grained features, while pooling layers reduce redundant information in feature maps; their combination can reduce computational complexity while maintaining feature richness. Separate convolutional and pooling layers can be flexibly combined into different network structures, allowing for optimization of the model architecture through experimentation with various configurations. Greater adaptability: Designing separate convolutional and pooling strategies for specific tasks and datasets can improve the model's adaptability and performance.

[0083] Understandably, automatically generating network models through algorithms can significantly reduce the workload of manual design, especially in scenarios with complex network structures or changing requirements. Automated algorithms can quickly search a large number of possible architecture combinations and select the optimal model structure, thereby improving model performance. Using identifier matrices and information flow matrices, the network structure is dynamically adjusted based on the characteristics of the input data, enabling the model to adapt to various task requirements. Automated construction methods can explore a wider network architecture design space, generating network combinations of different levels and types to meet complex application scenarios. The automated construction process facilitates standardized operations and can be repeatedly applied to different tasks and datasets, significantly improving development efficiency.

[0084] In this embodiment, the neural network structure is converted into a tree data structure, and the tree data structure is optimized using a depth-first search algorithm to generate an executable neural network model, including: By traversing the layers of the neural network structure, the nodes of the neural network structure are defined as nodes in a tree data structure, and each node is represented as a triple. The root node is defined as a node with no input operations, and the leaf nodes are defined as terminal nodes in the tree data structure with no child nodes. The value of each leaf node contains the set of operators and neural network nodes corresponding to that leaf node, thus constructing the initial tree data structure. The value of the node represents the state of the neuron; For the index of the node; It is a set of child nodes; Starting from the root node, recursively traverse all child nodes, recording the set of all paths from the root node to a leaf node. :

[0085] in, For path i The node values ​​in The length of the path; For path set Add dimensionality reduction layers and linear layers to obtain an executable neural network model.

[0086] The recursive call description of DFS is as follows:

[0087] In the formula, This indicates the operation of finding the current node and its child nodes.

[0088] By adding dimensionality reduction layers and linear layers to the optimized tree data structure, an executable neural network model is obtained.

[0089] like Figure 3 and Figure 4 As shown, Figure 3 This is a schematic diagram illustrating the process of constructing a tree using operators and neural network nodes according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating the execution of the depth-first search algorithm according to an embodiment of the present invention. Figure 3 In this tree, the root node (None, -1) represents the original input, initialized with an empty operator. Each branch node represents a different operator (e.g., e, a, b, c, d) and its associated neural network node (e.g., 1, 2, 3). Operators are added to the tree sequentially according to the connection relationships in the identifier matrix, forming a complete path from the root node to a leaf node. Figure 4 In this context, the process of traversing a tree structure using the Depth-First Search (DFS) algorithm begins at the root node (e, 0), indicating that data transmission starts from the input node. The algorithm visits all branches along the tree in a depth-first manner, exploring each path in advance until it reaches a leaf node (e.g., c, 3). The path from the root node to the leaf node records the complete connection relationships between neural network nodes and operators, ultimately forming a directly trainable network model.

[0090] In this embodiment, a loss function is constructed, and the executable neural network model is trained using an optimizer and a learning rate scheduler, including: Using the smooth L1 loss function as the training loss function, its expression is as follows:

[0091] In the formula, , For the true value, The predicted value of the model; if there is For each sample, the overall expression of the loss function is:

[0092] The Adam optimizer is used to update the training parameters of the executable neural network model, including: Calculate the current value according to the following formula. t First moment estimation at time 1 , representing the exponentially weighted average of the gradient:

[0093] in, Let be the exponentially weighted decay rate of the first moment, set to 0.9. For the current gradient, for t- First-order moment estimate at time 1; Calculate the current value according to the following formula. t Second moment estimation at time 1 , representing the exponentially weighted average of the squared gradient:

[0094] in, The exponentially weighted decay rate of the second moment is set to 0.9; The first moment is estimated according to the following formula. and estimation of second moment Perform deviation correction processing:

[0095]

[0096] in, This is the deviation correction value for the first moment. This is the deviation correction value for the second moment. for of t Power of 1 for of t Power of; The training parameters of the executable neural network model are updated according to the following formula:

[0097] in, For the first t The training parameter values ​​after the next iteration For the first The training parameter values ​​at the next iteration For learning rate, To prevent division by zero, let the constant be set as ; The cosine annealing learning rate scheduler is used to update the learning rate of an executable neural network model, as shown in the following expression:

[0098] in, Let be the learning rate for cosine annealing, and let represent the learning rate at the th... t The learning rate used during each training session. To minimize the learning rate, The initial learning rate, For training cycles, This is the current training round.

[0099] The mean squared error loss function is chosen as the loss function for the test set to measure the model's accuracy. Its definition is as follows:

[0100] In the formula, For the true value, The predicted value of the model. This represents the number of samples.

[0101] After 250 training rounds, the loss and performance metrics are calculated on the training and test sets respectively, and train_loss, test_loss, mse, and r are saved to a dictionary, where r is the Pearson coefficient.

[0102] Understandably, using a cosine function scheduler to dynamically adjust the learning rate, gradually decreasing it as training progresses, makes the optimization process more refined. By periodically restarting the learning rate, the model explores near local optima, helping it quickly find the global optimum. This learning rate restart mechanism prevents the model from stopping training due to an excessively low learning rate, ensuring a more thorough optimization process. The scheduler supports multiple restarts, allowing the model to optimize with different strategies at different training stages, thus adapting to complex loss surfaces.

[0103] like Figure 5 As shown, Figure 5 This is a schematic diagram illustrating the change of the loss function during the training process according to an embodiment of the present invention. Figure 5 This diagram illustrates the changes in training and validation losses over training epochs. The training loss (blue curve) decreases rapidly in the initial stages, indicating a gradual improvement in the model's fit on the training set. It then stabilizes and remains at a low level, demonstrating good convergence. The validation loss (orange curve) shows relatively small fluctuations overall. Although there are some fluctuations, it remains stable and within a low range, reflecting the model's good generalization ability and robustness. The small difference between the training and validation losses indicates that the model has not experienced overfitting. Overall, with the combination of an optimizer (such as Adam) and a learning rate scheduler (such as cosine annealing), the model effectively learns features and achieves performance optimization. In this embodiment, after the executable neural network model has been trained but before the meme algorithm is used to optimize the trained neural network model, the method further includes: The Pearson coefficient of the trained neural network model is calculated as the initial fitness value for the meme algorithm according to the following formula:

[0104] in, For the number of samples, the variable values These are the predicted values ​​of the neural network model, representing variable values. This represents the true breeding phenotypic label value. for and The sum of the products of corresponding values For variable values The sum, For variable values The sum, For variable values The sum of squares, For variable values The sum of squares.

[0105] In this embodiment, as Figure 6 As shown, Figure 6 This is a flowchart illustrating the execution of the meme algorithm in an embodiment of the present invention. The meme algorithm is used to optimize the trained neural network model to obtain the optimal breeding phenotype prediction model. The specific steps include: Step a: Using two trained neural network models as individuals in the population, select individuals from the population using the roulette wheel selection algorithm based on the initial fitness values.

[0106] The process involves selecting individuals from the population using a roulette wheel selection algorithm based on their initial fitness values, including: Calculate the sum of fitness values ​​for all individuals using the following formula. :

[0107] in, Indicates the first Individual, Indicates calculation fitness value; Calculate the individual according to the following formula. Probability of being selected :

[0108] Construct the cumulative probability distribution according to the following formula. :

[0109] In the formula, Indicates the selection of individuals The cumulative probability, Generate random numbers To obtain satisfaction individual As the chosen individual.

[0110] Step b: Perform a crossover operation on the selected individuals to generate new candidate individuals.

[0111] As an example, the two individuals selected will be designated as parent individuals. and parental individuals Choose an intersection point and alternately swap parent individuals. and parental individuals The gene segment is used to generate two offspring individuals as new candidate individuals, as shown in the following expression:

[0112]

[0113] in, and From the paternal generation and The offspring solutions generated after crossover.

[0114] Step c: Perform mutation operations on the new candidate individuals.

[0115] As an example, mutation operations on new candidate individuals can be performed in two ways: The first method involves randomly selecting a gene locus from a new candidate individual and changing the value of that gene locus to another value.

[0116] As an example, assume an individual The gene is represented as At the site Mutation occurs, and the resulting individual... for:

[0117] in, This is a newly generated value.

[0118] The second method involves randomly selecting two gene loci for exchange between two new candidate individuals.

[0119] As an example, two gene loci were randomly selected. and Exchange their values. If individuals mutated individuals for:

[0120]

[0121] In the formula, The remaining sites remain unchanged.

[0122] Step d: Perform a local search on the individual after the mutation operation and recalculate the individual's fitness value.

[0123] For each individual after mutation, a separation operation is performed on the neural network model to obtain a convolutional information stream containing only convolution operators and a pooling information stream containing only pooling operators.

[0124] As an example, the separation operation is defined as follows:

[0125] In the formula, It is a collection of convolutional layers. It is a collection of pooling layers.

[0126] The iteration count is set to itro. The mutation operation, as described in this embodiment, is performed on each pooling information stream to obtain a new pooling information stream. The separated convolutional information stream and the new pooling information stream are then merged, and the new neural network model replaces the current neural network model. The merging operation between information streams is defined as follows:

[0127] Here, error indicates that an error is thrown. Represents a separate convolutional information stream. This represents the pooled information flow after the mutation. The neural network model is formed by merging the separate convolutional information stream and the mutated pooling information stream.

[0128] New individuals are created using a novel neural network model, and their fitness values ​​are calculated.

[0129] Step e: Iteratively select the individual with the highest fitness value using the elite selection algorithm and the tournament selection algorithm, and use the neural network model of the individual with the highest fitness value as the optimal breeding phenotype prediction model.

[0130] As an example, the elite selection process is as follows: Construct set ,in i An index for an individual.

[0131] Based on fitness score Sort set S from highest to lowest to obtain a new set. .

[0132] Select elite individuals according to the following formula:

[0133] According to the following formula, for the set of elite individuals Perform column splitting:

[0134] in, For the current population, For fitness score, For model set, This is a set of node matrices. Define a function. ,in It is a collection of individuals selected by the elite.

[0135] As an example, the tournament selection process is as follows: Individuals for the tournament are randomly selected from the current population P according to the following formula:

[0136] In the formula, It is a random function. It is the first Individual indexes selected in the next tournament.

[0137] make ,in It is the first A subset of individuals in the tournament.

[0138] In each tournament, the individual with the highest fitness is selected according to the following formula. :

[0139] The winners of each tournament will be grouped together to form a team. This leads to the formation of new populations.

[0140] in, For the current population, For fitness score, For model set, Let the set be a matrix of nodes. Define a random selection function. ,in It is the set of indices for the selected individuals.

[0141] Understandably, meme algorithms balance efficiency and accuracy in global and local searches by adjusting the search strategy according to the objective function. Meme algorithms emphasize maintaining population diversity to avoid premature convergence, thus allowing for a broader exploration of the search space. They can escape local optima and find better global solutions through diverse population search strategies. Building upon global search, local mutation and optimization further improve the accuracy and generalization ability of the solution. When optimizing complex network structures, meme algorithms can more effectively find model parameters with higher fitness, thereby improving the overall predictive performance of the model.

[0142] Table 1 Comparison of breeding phenotypic prediction accuracy of different models

[0143] Table 1 shows the comparison of prediction accuracy of different models in two environments: TKW (plant height) and GP (particle hardness). The MA model of this invention performed best, achieving Pearson correlation coefficients of 0.69 and 0.52, respectively, indicating its superior overall performance in predicting complex traits. In contrast, the traditional method GBLUP had the lowest accuracy, with correlation coefficients of 0.18 and 0.12 in the TKW and GP environments, respectively, indicating its difficulty in effectively handling large-scale gene data. Traditional machine learning models LightGBM and SVR performed moderately; although LightGBM achieved 0.64 in the TKW environment, it failed to surpass the MA model. Deep learning-based models, such as DeepGS and DNNGP, performed well in both environments. DNNGP achieved accuracies of 0.67 and 0.5, respectively, approaching the performance of the MA model but still slightly inferior. DLGWAS performed relatively stably in both environments (0.63 and 0.48), but also failed to surpass the MA model. Overall, the MA model combines global search and local optimization capabilities, demonstrating significant predictive performance advantages and validating its application value and reliability in modern breeding research.

[0144] Understandably, this invention proposes a novel breeding phenotypic prediction model that, by introducing a multi-layered and complex structure, aims to effectively address several shortcomings of traditional breeding phenotypic prediction methods. The model employs a deep learning architecture, capable of capturing the non-linear relationship between traits and genotypes, overcoming the limitation of linear mixture models in failing to identify complex interaction effects. By introducing advanced feature extraction mechanisms, pooling layers, and attention mechanisms, the prediction accuracy of complex quantitative traits (such as feed conversion ratio, yield, and disease resistance) is significantly improved. Simultaneously, the model utilizes multi-channel input and a flexible structural design, exhibiting stronger robustness to variations in input data and the ability to handle diverse biological data. Furthermore, the depth and complexity of the model structure endow it with good scalability, making it suitable for different scenarios and surpassing the limitations of simple models in specific tasks. This comprehensive approach not only improves predictive performance but also provides a better biological explanation of the relationship between genes and traits, demonstrating significant application value in modern breeding and genomics research.

[0145] This invention utilizes multi-layer convolution and deep learning mechanisms to automatically extract key features from high-dimensional, complex data. Combined with the MA intelligent algorithm to optimize the model structure, it significantly improves the efficiency and accuracy of feature learning. Simultaneously, the model effectively captures the complex nonlinear relationships between genotypes and traits using nonlinear activation functions and ensures data shape consistency through a forward propagation mechanism, thereby enhancing robustness to input data. In predicting complex traits (such as feed conversion ratio, yield, and disease resistance), the model demonstrates higher prediction accuracy by combining the DFS (Depth-First Search) algorithm with comprehensive consideration of multiple genes and environmental factors. Furthermore, the model exhibits strong flexibility and adaptability. Through decoupling and stacking of operators and neural network nodes, it can be adjusted and optimized according to task requirements, making it suitable for various genomics research and breeding applications. This flexibility not only promotes scientific decision-making in modern agricultural breeding and genetic research but also drives the development of bioinformatics and biotechnology, providing strong technical support for precision agriculture and personalized breeding.

[0146] Example 3 like Figure 7 As shown, this embodiment proposes a breeding phenotype prediction system for gene sequence data, which is applied to the breeding phenotype prediction method for gene sequence data as described in the above embodiment. It includes: an acquisition module 100, a partitioning module 200, a construction module 300, an allocation module 400, a conversion optimization module 500, a training optimization module 600, and a prediction module 700.

[0147] The system comprises the following modules: an acquisition module 100 acquires genotype and phenotypic data and constructs a dataset using these data; a partitioning module 200 partitions the dataset into a training set and a test set; a construction module 300 constructs a convolution operator space, a pooling operator space, and neural network nodes; an allocation module 400 allocates convolution and pooling operators to the neural network nodes from the convolution and pooling operator spaces to generate a neural network structure; a conversion and optimization module 500 converts the neural network structure into a tree-like data structure and optimizes the tree-like data structure using a depth-first search algorithm to generate an executable neural network model; a training and optimization module 600 trains the executable neural network model using the training set and optimizes the trained neural network model using a meme algorithm to obtain the optimal breeding phenotypic prediction model; and a prediction module 700 inputs the test set data into the optimal breeding phenotypic prediction model for processing to obtain the breeding phenotypic prediction result.

[0148] It should be noted that the explanation of the above-mentioned breeding phenotype prediction method embodiment for gene sequence data also applies to the breeding phenotype prediction system for gene sequence data in this embodiment, and will not be repeated here.

[0149] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0150] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0151] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more N executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of the invention pertain.

[0152] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0153] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium. When executed, the program includes one or a combination of the steps of the method embodiments.

[0154] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A breeding phenotype prediction method based on gene sequence data, characterized in that, include: Acquire genotype and phenotypic data, and construct a dataset using the genotype and phenotypic data; The dataset is divided into a training set and a test set; Construct the convolution operator space, pooling operator space, and neural network nodes; From the convolution operator space and the pooling operator space, convolution operators and pooling operators are assigned to neural network nodes to generate a neural network structure; The neural network structure is converted into a tree data structure, and the tree data structure is optimized using a depth-first search algorithm to generate an executable neural network model, including: By traversing the layers of the neural network structure, the nodes of the neural network structure are defined as nodes in a tree data structure, and each node is represented as a triple. The root node is defined as a node with no input operations, and the leaf nodes are defined as terminal nodes in the tree data structure with no child nodes. The value of each leaf node contains the set of operators and neural network nodes corresponding to that leaf node, thus constructing the initial tree data structure. The value of the node represents the state of the neuron; For the index of the node; It is a set of child nodes; Starting from the root node, recursively traverse all child nodes, recording the set of all paths from the root node to a leaf node. : in, For path i The node values ​​in The length of the path; For path set Adding dimensionality reduction layers and linear layers yields an executable neural network model; The executable neural network model is trained using the training set, and the trained neural network model is optimized using the meme algorithm to obtain the optimal breeding phenotype prediction model. The test set data is input into the optimal breeding phenotype prediction model for processing to obtain the breeding phenotype prediction results.

2. The breeding phenotype prediction method based on gene sequence data according to claim 1, characterized in that, After dividing the dataset into training and test sets, the method further includes: For even-dimensional feature data in the training set, the encoding is performed according to the following formula: For odd-dimensional feature data in the training set, the encoding is performed according to the following formula: in, Indicates the location index of the sample data. The dimension index of the feature vector. Indicates the dimension of the feature vector.

3. The breeding phenotype prediction method based on gene sequence data according to claim 1, characterized in that, Construct the convolution operator space and pooling operator space according to the following formula: in, It represents a set of characters, where each character is used to identify a convolution operator or a pooling operator; For the set of convolution operators, characters Different convolution operators, characters This indicates no connection or no operation. For the pooling operator set, characters Different pooling operators, To store the set of operators selected from the convolution operator space or the pooling operator space, For the convolution operator space, For the pooling operator space, This indicates that characters are mapped to the corresponding convolution operators or pooling operations.

4. The breeding phenotype prediction method based on gene sequence data according to claim 1, characterized in that, The neural network node includes a CBAM attention mechanism and multiple convolutional blocks with residual connections; The CBAM attention mechanism includes a channel attention mechanism and a spatial attention mechanism; The channel attention mechanism is used to weight the different channels of the input data, including: The channel attention weights are calculated according to the following formula. : in, The feature map representing the input data, For the number of channels, and The height and width of the feature map; and For different learning weight matrices, For the reduction rate, For activation function, This represents the Sigmoid activation function; Generate the channel attention weight feature map according to the following formula. : in, This represents element-wise dot product; The spatial attention mechanism is used to weight the spatial location of the input data, including: The spatial attention weights are calculated according to the following formula. : in, Indicates the feature map Average pooling is performed along the channel dimension. This represents max pooling of the feature maps along the channel dimension, while concat() represents concatenating two feature maps along the channel dimension. Generate a spatial attention map according to the following formula. : The output of each neural network node is generated according to the following formula: in, Indicates the first The nth neural network node is related to the original input or the nth neural network node. The connection method of each neural network node This refers to the index position information of the neural network node, used to determine the first... Each neural network node and Is there a connection between them, characters? This indicates no connection or no operation. This represents the computational operation unit of the current neural network node.

5. The breeding phenotype prediction method based on gene sequence data according to claim 1, characterized in that, Constructing a loss function and training the executable neural network model using an optimizer and a learning rate scheduler includes: Using the smooth L1 loss function as the training loss function, its expression is as follows: In the formula, , For the true value, These are the model's predicted values; The Adam optimizer is used to update the training parameters of the executable neural network model, including: Calculate the current value according to the following formula. t First moment estimation at time 1 , representing the exponentially weighted average of the gradient: in, The exponentially weighted decay rate of the first moment. For the current gradient, for t- First-order moment estimate at time 1; Calculate the current value according to the following formula. t Second moment estimation at time 1 , representing the exponentially weighted average of the squared gradient: in, The exponentially weighted decay rate of the second moment; The first moment is estimated according to the following formula. and estimation of second moment Perform deviation correction processing: in, This is the deviation correction value for the first moment. This is the deviation correction value for the second moment. for of t Power of 1 for of t Power of; The training parameters of the executable neural network model are updated according to the following formula: in, For the first t The training parameter values ​​after the next iteration For the first The training parameter values ​​at the next iteration For learning rate, To prevent division by zero of constants; The cosine annealing learning rate scheduler is used to update the learning rate of an executable neural network model, as shown in the following expression: in, Let be the learning rate for cosine annealing, and let represent the learning rate at the th... t The learning rate used during each training session. To minimize the learning rate, The initial learning rate, For training cycles, This is the current training round.

6. The breeding phenotype prediction method based on gene sequence data according to claim 1, characterized in that, After training the executable neural network model and before optimizing the trained neural network model using a meme algorithm, the method further includes: The Pearson coefficient of the trained neural network model is calculated as the initial fitness value for the meme algorithm according to the following formula: in, For the number of samples, the variable values These are the predicted values ​​of the neural network model, representing variable values. This represents the true breeding phenotypic label value. for and The sum of the products of corresponding values For variable values The sum, For variable values The sum, For variable values The sum of squares, For variable values The sum of squares.

7. The breeding phenotype prediction method based on gene sequence data according to claim 6, characterized in that, The trained neural network model is optimized using a meme algorithm to obtain the optimal breeding phenotype prediction model, including: Using two trained neural network models as individuals in the population, individuals are selected from the population using a roulette wheel selection algorithm based on their initial fitness values. Perform crossover operations on the selected individuals to generate new candidate individuals; Perform mutation operations on new candidate individuals; Perform a local search on the individuals after mutation and recalculate their fitness values; The individual with the highest fitness value is selected iteratively using the elite selection algorithm and the tournament selection algorithm. The neural network model of the individual with the highest fitness value is then used as the optimal breeding phenotype prediction model.

8. The breeding phenotype prediction method based on gene sequence data according to claim 7, characterized in that, Based on the initial fitness values, individuals are selected from the population using the roulette wheel selection algorithm, including: Calculate the sum of fitness values ​​for all individuals using the following formula. : in, Indicates the first Individual, Indicates calculation fitness value; Calculate the individual according to the following formula. Probability of being selected : Construct the cumulative probability distribution according to the following formula. : In the formula, Indicates the selection of individuals The cumulative probability, Generate random numbers To obtain satisfaction individual As the selected individual; Perform crossover operations on the selected individuals to generate new candidate individuals, including: The two individuals selected will be designated as parent individuals. and parental individuals ; Select the intersection point and alternately swap the parent individuals. and parental individuals The gene segment is used to generate two offspring individuals as new candidate individuals; Mutation operations are performed on new candidate individuals, including: Randomly select a gene locus from a new candidate individual and change the value of that gene locus to another value, or randomly select two gene loci between two new candidate individuals and exchange them; Perform a local search on the individuals after mutation, and recalculate the fitness values ​​of the individuals, including: For each individual after mutation, a separation operation is performed on the neural network model to obtain a convolutional information stream containing only convolution operators and a pooling information stream containing only pooling operators; The pooling information stream is iteratively mutated, and the mutated pooling information stream and convolutional information stream are merged to obtain a new neural network model. New individuals are created using a novel neural network model, and their fitness values ​​are calculated.

9. A breeding phenotype prediction system based on gene sequence data, applied to the breeding phenotype prediction method based on gene sequence data as described in any one of claims 1 to 8, characterized in that, include: The acquisition module is used to acquire genotype data and phenotypic data, and to construct a dataset using the genotype data and phenotypic data; The dataset is divided into training and testing sets by partitioning it into modules. The building blocks are used to construct the convolution operator space, pooling operator space, and neural network nodes; The allocation module is used to allocate convolution operators and pooling operators to neural network nodes from the convolution operator space and the pooling operator space to generate a neural network structure; The conversion and optimization module is used to convert the neural network structure into a tree data structure and optimize the tree data structure using a depth-first search algorithm to generate an executable neural network model. The training optimization module is used to train the executable neural network model using the training set, and to optimize the trained neural network model using the meme algorithm to obtain the optimal breeding phenotype prediction model. The prediction module is used to input the test set data into the optimal breeding phenotype prediction model for processing, and obtain the breeding phenotype prediction results.

Images