Breeding phenotype prediction method and system oriented to gene sequence data

Abstract

The invention relates to the technical field of breeding phenotype prediction, and discloses a breeding phenotype prediction method and system oriented to gene sequence data, and the method comprises the steps: obtaining 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 neural network nodes, distributing convolution operators and pooling operators for the neural network nodes, and generating a neural network structure; the neural network structure is converted into a tree data structure, a depth-first search algorithm is used for optimization, and an executable neural network model is generated; training an executable neural network model by using the training set, and optimizing the trained neural network model by using a memetic 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. According to the method, the prediction precision is remarkably improved in the breeding phenotype prediction of complex characters.

Description

Technical Field

[0001] The present invention relates to the technical field of breeding phenotype prediction, and more specifically, to a breeding phenotype prediction method and system for gene sequence data. Background Art

[0002] With the continuous growth of the global population and the changing environmental conditions, the issue of food security has become increasingly severe. Modern breeding technologies improve the genetic characteristics of species, enhancing their yield, disease resistance, nutritional quality, and environmental adaptability, providing an important way to address food shortages. The rapid development of precision breeding and gene editing technologies has made it possible to explore genetic diversity and introduce excellent traits at the genomic level. However, the genetic improvement of complex traits (such as feed conversion ratio, stress resistance, etc.) 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] Existing technologies mainly adopt the Genomic Selection (GS) method to predict the genetic potential of individuals 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 through high-throughput sequencing data and deep neural network models, can capture the complex relationship between genotype and phenotype, and improve the efficiency and accuracy of complex trait prediction. In addition, some studies combine the combination of convolutional neural network (CNN) and fully connected layers to handle the multi-gene and genotype-by-environment interaction effects (GxE), further enhancing the prediction ability for complex traits.

[0004] However, traditional GS methods are difficult to capture the non-linear relationship between genotype and traits, and insufficient consideration of gene-environment interaction effects leads to low prediction accuracy for complex traits. At the same time, deep learning-based models (such as the DNNGP method) also face many problems in practical applications: the model lacks the ability to process multi-channel data input, has insufficient robustness to the shape of input data, limited high-level feature extraction ability, and the network structure is too simple to adapt to complex breeding needs. These defects directly limit the prediction accuracy of complex traits. Summary of the Invention

[0005] In order to improve the prediction accuracy of breeding phenotype prediction technology, the present invention proposes the following technical solutions: In a first aspect, the present invention proposes a breeding phenotype prediction method for gene sequence data, including: Obtain genotype data and phenotype data, and construct a data set using the genotype data and the phenotype data; Divide the data set into a training set and a test set; Construct a convolutional operator space, a pooling operator space, and neural network nodes; Allocate a convolutional operator and a pooling operator for the neural network nodes from the convolutional operator space and the pooling operator space to generate a neural network structure; 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; Use a training set to train the executable neural network model, and optimize the trained neural network model using a memetic algorithm to obtain an optimal breeding phenotype prediction model; Input the test set data into the optimal breeding phenotype prediction model for processing to obtain a breeding phenotype prediction result.

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

[0007] For the odd-dimensional feature data in the training set, encode it according to the following formula:

[0008] Where, represents the position index of the sample data, represents the dimension index of the feature vector, represents the dimension of the feature vector.

[0009] As a preferred technical solution, construct a convolutional operator space and a pooling operator space according to the following formula:

[0010] Where, represents a character set, and each character in the set is used to identify a convolutional operator or a pooling operator; is a set of convolutional operators, and the characters are different convolutional operators respectively, and the character represents no connection or no operation; is a set of pooling operators, and the characters are different pooling operators respectively, is a set for storing the operators selected from the convolutional operator space or the pooling operator space, is the convolutional operator space, is the pooling operator space, represents mapping a character to the corresponding convolutional operator or pooling.

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

[0012] where, represents the feature map of the input data, is the number of channels, and are the height and width of the feature map; and are different learned weight matrices, is the reduction rate, is the activation function, Sigmoid activation function; Generate a channel attention weight feature map according to the following formula :

[0013] where, represents element-wise multiplication; The spatial attention mechanism is used to weight the weights of the spatial positions of the input data, including: Calculate the spatial attention weight according to the following formula :

[0014] where, represents average pooling of the feature map in the channel dimension, represents max pooling of the feature map in the channel dimension, and concat() represents concatenating two feature maps in 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] where, represents the th neural network node and the original input or the The connection method of a neural network node is the index position information of the neural network node, used to determine whether there is a connection between the th neural network node and ; 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 levels of the neural network structure, the nodes of the neural network structure are defined as nodes in the tree data structure. The nodes are represented as triples . 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 that have no child nodes. Moreover, the values of the leaf nodes contain the set of operators and neural network nodes corresponding to the leaf nodes, thus constructing an initial tree data structure; where is the value of the node, representing the state of the neuron; is the index of the node; is the set of child nodes; Starting from the root node, recursively traverse all child nodes and record the set of all paths from the root node to the leaf nodes :

[0018] where is the node value in the path i , is the length of the path; Add a dimensionality reduction layer and a linear layer to the path set to obtain an executable neural network model.

[0019] As a preferred technical solution, construct a loss function, and use an optimizer and a learning rate scheduler to train the executable neural network model, including:[[]] Use the smooth L1 loss function as the loss function for training, and its expression is as follows:[[]]

[0020] In the formula, , is the true value, is the predicted value of the model; Use the Adam optimizer to update the training parameters of the executable neural network model, including:[[]] According to the following formula, calculate the first-order moment estimate t at the current , representing the exponentially weighted average of the gradient:

[0021] where, is the exponentially weighted decay rate of the first moment, is the current gradient, is t- the estimate of the first moment at time 1; According to the following formula, calculate the estimate of the second moment at the current t time , representing the exponentially weighted average of the squared gradient:

[0022] where, is the exponentially weighted decay rate of the second moment; According to the following formula, perform bias correction on the estimate of the first moment and the estimate of the second moment :

[0023]

[0024] where, is the bias correction value of the first moment, is the bias correction value of the second moment, is to the t power, is to the t power; According to the following formula, update the training parameters of the executable neural network model:

[0025] where, is the value of the training parameter after the t th iteration, is the value of the training parameter at the th iteration, is the learning rate, is a constant to prevent division by zero; Use the cosine annealing learning rate scheduler to update the learning rate of the training of the executable neural network model, and its expression is as follows:

[0026] where, is the learning rate of cosine annealing, representing the learning rate used at the t th training, is the minimum learning rate, is the initial learning rate, is the number of training epochs, is the current training iteration.

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

[0028] where, is the number of samples, and the variable value is the predicted value of the neural network model, and the variable value is the true breeding phenotypic label value, is and the sum of the products of the corresponding values, is the sum of the variable values , is the sum of the variable values , is the sum of the squares of the variable values , is the sum of the squares of the variable values .

[0029] As a preferred technical solution, using the memetic algorithm to optimize the trained neural network model to obtain the optimal breeding phenotypic prediction model includes: Taking two trained neural network models as individuals of the population, and selecting individuals from the population using the roulette wheel selection algorithm according to the initial fitness value; Performing a crossover operation on the selected individuals to generate new candidate individuals; Performing a mutation operation on the new candidate individuals; Performing a local search on the individuals after the mutation operation and recalculating the fitness values of the individuals; Iteratively selecting the individual with the highest fitness value through the elite selection algorithm and the tournament selection algorithm, and using the neural network model in the individual with the highest fitness as the optimal breeding phenotypic prediction model.

[0030] As a preferred technical solution, selecting individuals from the population using the roulette wheel selection algorithm according to the initial fitness value includes: Calculating the sum of the fitness values of all individuals according to the following formula :

[0031] Among them, represents the th individual, represents the calculation of fitness value; According to the following formula, calculate the probability that the individual is selected:

[0032] According to the following formula, construct the cumulative probability distribution :

[0033] In the formula, represents the cumulative probability of selecting the individual , generate a random number , and obtain the individual that satisfies as the selected individual; Perform crossover operation on the selected individuals to generate new candidate individuals, including: Take the two selected individuals as the parental individuals and parental individual respectively; Select the crossover point and alternately exchange the gene segments of the parental individual and parental individual to generate two offspring individuals as new candidate individuals; Perform mutation operation on the new candidate individuals, including: Randomly select a gene locus in the new candidate individuals and change the value of this gene locus to other values, or randomly select two gene loci between the two new candidate individuals and exchange them; Perform local search on the individuals after the mutation operation and recalculate the fitness value of the individuals, including: For each neural network model in the individuals after the mutation operation, perform a separation operation to obtain a convolution information flow containing only convolution operators and a pooling information flow containing only pooling operators; Perform iterative mutation operation on the pooling information flow and merge the mutated pooling information flow and the convolution information flow to obtain a new neural network model; Use the new neural network model to form a new individual and calculate the fitness value of the new individual.

[0034] In the second aspect, the present invention also proposes a breeding phenotype prediction system for gene sequence data, which is applied to the breeding phenotype prediction method for gene sequence data described in any one of the solutions in the first aspect, including: An acquisition module, configured to acquire genotype data and phenotype data, and construct a data set by using the genotype data and the phenotype data; A partitioning module, configured to partition the data set into a training set and a test set; A construction module, configured to construct a convolutional operator space, a pooling operator space, and neural network nodes; An allocation module, configured to allocate a convolutional operator and a pooling operator for the neural network nodes from the convolutional operator space and the pooling operator space, and generate a neural network structure; A conversion and optimization module, configured to convert the neural network structure into a tree data structure, and optimize the tree data structure by using a depth-first search algorithm, so as to generate an executable neural network model; A training and optimization module, configured to use the training set to train the executable neural network model, and optimize the trained neural network model by using a memetic algorithm, so as to obtain an optimal breeding phenotype prediction model; A prediction module, configured to input the test set data into the optimal breeding phenotype prediction model for processing, so as to obtain a breeding phenotype prediction result.

[0035] The beneficial effects of the present invention at least include: By constructing a convolutional operator space and a pooling operator space, the present invention dynamically allocates an adapted convolutional operator and a pooling operator for neural network nodes, generates an efficient and highly adaptable neural network structure, and optimizes the tree data structure of the neural network by using a depth-first search (DFS) algorithm to ensure the accuracy and efficiency of the feature extraction path. By combining a memetic algorithm to globally and locally optimize the network model, not only the convergence speed of the model is improved, but also the prediction ability for complex traits (such as feed conversion ratio, yield, disease resistance, etc.) is significantly enhanced. Through optimized design, the model can extract more biologically meaningful non-linear relationship features from high-dimensional genotype and phenotype data, effectively capture the complexity of multi-gene and environment interaction effects (GxE), and provide strong support for the accurate prediction of complex traits.

[0036] Meanwhile, by dynamically adjusting the connection structure of neural network nodes, the present invention ensures the consistency between the input data and the network structure, enhances the robustness of the model to changes in input data, and avoids prediction errors caused by data anomalies or shape mismatches. The model adopts an advanced feature extraction mechanism combining multi-layer convolution and pooling, effectively captures key biological signals in the input data, and significantly improves the extraction efficiency and accuracy of features required for complex trait prediction. Through optimization on the training set and verification on the test set, the model exhibits good adaptability and scalability, can be flexibly applied to different scenarios and data sets, and ensures the universality and stability of complex trait prediction.

[0037] Finally, the present invention significantly improves the prediction accuracy in the breeding phenotype prediction of complex traits (such as feed conversion ratio, yield, disease resistance, etc.), overcomes the technical bottlenecks of the traditional genomic selection method in capturing insufficient complex interaction effects and having relatively low prediction accuracy, and provides a reliable technical guarantee for modern breeding and genomics research. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic flowchart of the breeding phenotype prediction method for gene sequence data provided by an embodiment of the present invention.

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

[0040] Figure 3 It is a schematic flowchart of the process of constructing a tree of operators and neural network nodes in an embodiment of the present invention.

[0041] Figure 4 It is a schematic flowchart of the process of executing the depth-first search algorithm in an embodiment of the present invention.

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

[0043] Figure 6 It is a schematic flowchart of the process of executing the memetic algorithm in an embodiment of the present invention.

[0044] Figure 7 It is a schematic flowchart of the breeding phenotype prediction system for gene sequence data provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] The following will describe the embodiments of the present invention 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 explaining the present invention and not for limiting the protection scope of the present invention.

[0046] It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Therefore, only the components related to the present invention are shown in the diagrams, rather than being drawn according to the number, shape, and size of the components in actual implementation. The types, quantities, and proportions of the components in actual implementation can be arbitrarily changed, and the component layout type may also be more complex.

[0047] In the following description, numerous details are explored to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present invention difficult to understand.

[0048] Embodiment 1 This embodiment proposes a breeding phenotype prediction method for gene sequence data, as Figure 1 shown, Figure 1 is a schematic flowchart of a breeding phenotype prediction method for gene sequence data provided by this embodiment. The method includes the following steps: S1: Obtain genotype data and phenotype data, and construct a data set using the genotype data and the phenotype data.

[0049] S2: Divide the data set into a training set and a test set; S3: Construct a convolutional operator space, a pooling operator space, and neural network nodes; S4: Assign a convolutional operator and a pooling operator to the neural network nodes from the convolutional 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: Use the training set to train the executable neural network model, and use a memetic algorithm to optimize the trained neural network model to obtain an optimal breeding phenotype prediction model; S7: Input the test set data into the optimal breeding phenotype prediction model for processing to obtain a breeding phenotype prediction result.

[0050] It can be understood that the present invention constructs a convolutional operator space and a pooling operator space, dynamically assigns suitable convolutional operators and pooling operators to neural network nodes to generate an efficient and highly adaptable neural network structure, and uses a depth-first search (DFS) algorithm to optimize the tree data structure of the neural network to ensure the accuracy and efficiency of the feature extraction path. Combining a memetic algorithm to optimize the network model globally and locally not only improves the convergence speed of the model but also significantly enhances the prediction ability for complex traits (such as feed conversion ratio, yield, disease resistance, etc.). Through optimized design, the model can extract more biologically meaningful non-linear relationship features from high-dimensional genotype and phenotype data, effectively capture the complexity of multi-gene and environment interaction effects (GxE), and provide strong support for the accurate prediction of complex traits.

[0051] Meanwhile, the present invention ensures the consistency between the input data and the network structure by dynamically adjusting the connection structure of the neural network nodes, enhancing the robustness of the model to changes in the input data and avoiding prediction errors caused by data anomalies or shape mismatches. The model adopts an advanced feature extraction mechanism combining multi-layer convolution and pooling to effectively capture the key biological signals in the input data, significantly improving the extraction efficiency and accuracy of the 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, being able to be flexibly applied to different scenarios and datasets, ensuring the generality and stability of complex trait prediction.

[0052] Finally, the present invention significantly improves the prediction accuracy in the breeding phenotype prediction of complex traits (such as feed conversion ratio, yield, disease resistance, etc.), overcomes the technical bottlenecks of traditional genomic selection methods in capturing insufficient complex interaction effects and having relatively low prediction accuracy, and provides a reliable technical guarantee for modern breeding and genomics research.

[0053] Example 2 This example makes improvements based on the breeding phenotype prediction method for gene sequence data proposed in Example 1.

[0054] In this example, genotype data and phenotype data are obtained from public databases, such as high-throughput sequencing data and field phenotype measurement data of crops like wheat and corn. In the data processing stage, the phenotype data (such as plant height, disease resistance, feed conversion ratio, etc.) is standardized to ensure that these data can fully reflect the actual breeding goals. The original data is reorganized 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 a training set and a test set according to a ratio of 0.9, where the training set has a shape of (1800, 1, 1691) and the test set has a shape of (200, 1, 1691). A positional encoding operation is performed on the training set, which fuses the sample position index with the feature vector dimension.

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

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

[0057] where represents the position index of the sample data, represents the dimension index of the feature vector, represents the dimension of the feature vector.

[0058] Wrap the encoded training set data with DataLoader. DataLoader has the following functions: supporting efficient data loading from memory or disk, batching data by batch, avoiding biases caused by data order during training, and supporting multi-threading to accelerate data reading.

[0059] It can be understood that by adding unique encodings to each position in the sequence, the model can understand the order and position information of the input data. This encoding not only helps the model capture the relative position information in the sequence but also improves the representation ability 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 distinguishability in the high-dimensional space, which helps the model distinguish inputs at different positions. The sine-cosine position encoding does not depend on the input sequence of a fixed length and can dynamically calculate the position encoding of any length as needed, making it applicable to various sequence data of different scales. This flexibility greatly improves the generality of the model.

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

[0061] where, represents the character set, and each character in the set is used to identify a convolution operator or a pooling operator; is the convolution operator set, and the characters are different convolution operators respectively, and the character represents no connection or no operation; is the pooling operator set, and the characters are different pooling operators respectively, is the set for storing the operators selected from the convolution operator space or the pooling operator space, is the convolution operator space, is the pooling operator space, represents mapping the character to the corresponding convolution operator or pooling.

[0062] It should be noted that the convolution operator can identify key gene fragments related to complex traits in genotype data. For example, when predicting wheat disease resistance, the convolution operator can extract potential patterns in the region of disease resistance gene loci. The pooling operator compresses redundant information by reducing the dimension of the feature map, enabling the model to maintain efficient operation when processing large-scale genotype data and capturing the macroscopic connection between phenotypes and genotypes. Each convolution operator is encapsulated as a class, and the convolution operation on the input data is implemented in the class. After encapsulation, it is convenient for modular calling and flexible adjustment. The pooling operator does not require encapsulation of a class and can be directly used, reducing the complexity of the operator space. In the convolution operator space and the pooling operator space, characters are mapped to operator objects through key-value pairs. Subsequently, the random method is used to randomly return characters and corresponding operators from the convolution operator space and the pooling operator space for generating the model structure. Neural network nodes are predefined as classes and are created cyclically according to the number of nodes, with a fixed convolution kernel size. As Figure 2 shown Figure 2 in, x Figure 4 is an example diagram of the connection and stacking of the operators and neural network nodes in the embodiment of the present invention. The input

[0063] is the initial data of the model, which is sequentially transmitted and processed through multiple neural network nodes and operators. Among them: the node numbers (such as 0, 1, 2, 3) represent neural network nodes. The operators (such as a, b, c, d, e) connect different neural network nodes and are used to perform operations such as convolution or pooling. The arrow direction represents the transmission path of the data stream. Through these paths, the data is gradually transmitted from the input node to the output node. In this embodiment, the neural network node includes a CBAN attention mechanism and multiple convolution blocks with residual connections; The CBAN attention mechanism includes a channel attention mechanism and a spatial attention mechanism; The channel attention mechanism is used to weight the weights of different channels of the input data, including:

[0064] where, represents the feature map of the input data; is the number of channels; and are the height and width of the feature map; and are different learned weight matrices; is the reduction rate, taking a value of 16 or 8; is the activation function, using ReLU; The Sigmoid activation function is used to compress the output to the range [0, 1].

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

[0066] where represents element-wise multiplication; The spatial attention mechanism is used to weight the weights of the spatial positions of the input data, including: Calculate the spatial attention weight according to the following formula , representing the importance of each spatial position:

[0067] where represents average pooling of the feature map in the channel dimension, and the result is 's feature map; represents max pooling of the feature map in the channel dimension, and the result is also 's feature map; concat() represents concatenating two feature maps in the channel dimension.

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

[0069] It can be understood that CBAM (Channel and Spatial Attention Module) can significantly improve the task performance by adaptively adjusting the channel weights and spatial weights, enabling the model to automatically focus on key feature regions. The CBAM module has a compact structure and a simple design, and can be easily integrated into existing convolutional neural networks without significantly modifying the network architecture. This feature makes CBAM an efficient and easy-to-use enhancement module. After introducing the attention mechanism, CBAM can intuitively show the channels and spatial regions that the network focuses on when processing input data, which helps to understand the decision-making basis and feature extraction process of the model.

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

[0071] For the th neural network node, it has a total of information flow positions for the convolution operator and the pooling operator to connect. Among them, represents the connection method between the current neural network node and the original input and the neural network node. If 0 represents the original input, then the The output of a neural network node is as follows:

[0072] Among them, represents the connection method between the -th neural network node and the original input or the -th neural network node, is the index position information of the neural network node, used to determine whether there is a connection between the -th neural network node and , represents the computational operation unit of the current neural network node.

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

[0074] Among them, represents the input feature map; usually the output of the previous layer, represents a learnable function, usually composed of multiple neural network layers, such as convolutional layers, activation functions (such as ReLU), and normalization layers (such as Batch Normalization); represents the output feature map after the residual connection, which contains the sum of the input feature map and the processed feature map.

[0075] It can be understood that in deep networks, gradients may gradually vanish during backpropagation, leading to difficult training. Residual connections significantly alleviate this problem by allowing gradients to directly propagate from the output layer to the input layer, thus improving the feasibility of training deep networks. Residual connections reduce the optimization difficulty by learning the residuals (i.e., the differences between the input and the target) rather than the complete mapping, enabling 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 enhancing the overall performance. Residual connections allow the features extracted by the previous layer to be directly reused by the subsequent layer, ensuring that important information is not lost. This feature reuse mechanism further enhances the expressive power of the model.

[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 reallocated, specifically including: Create an empty two-dimensional array with a shape of (Neural_nodes, Neural_nodes + 1). 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 status of operators. Decode the neural network structure in a loop. The final decoded result is a set of characters (such as the character identifiers of convolutional operators or pooling operators) that record the internal information flow connections of each neural network structure.

[0077] Then, after decoding the initial model, further generate an identification matrix, whose purpose is to label the operator types of each connection in the network structure to ensure the execution and connection validity of the model. Each row of the identification matrix corresponds to a node, and each column is specifically defined as follows: -1 represents a pooling operation, 1 represents a convolutional operation, and 0 represents no connection. According to the identification matrix, generate the connection information of each node and determine which nodes have convolutional or pooling connections. This ensures that it can be identified which connections are valid, which are non-connections, and which are the results of convolutional or pooling operations. At the same time, combining the decoding result and the identification matrix, each set of neural network structures can be returned, clarifying the relationships between nodes and the network topology of the overall model.

[0078] Dynamically construct a layer of the neural network from the neural network structure and the identification matrix to avoid frequent memory allocation. Use an outer loop to control the number of times the network is constructed, traverse each identification matrix, and store the column information of each row in the ModuleList. Determine how to process each position according to the values (0, -1, 1, 2) in the identification matrix: 0 indicates no layer, and add a NoPlayer instance.

[0079] -1 represents a pooling operation, obtain and add it from the existing network layers.

[0080] 1 and 2 represent convolutional operations and neural network node operations. If the number of channels in the current layer does not meet the conditions, create a new layer to ensure that the model adapts to the data stream.

[0081] Finally, a network structure composed of multiple layers of neural networks is constructed. The configuration of each layer is flexibly adjusted according to the values in the matrix, ensuring the scalability and performance of the model.

[0082] It can be understood that the convolution operator focuses on feature extraction, learning local features and patterns (such as edges and textures), while the pooling operator reduces the spatial dimension of the feature map, reduces the computational cost and retains important information. After separation, the parameters of the convolutional layer and the pooling layer can be flexibly adjusted, such as the convolution kernel size, stride, and pooling method, etc., so as to adapt to various task requirements. The convolutional layer captures fine-grained features, and the pooling layer reduces the redundant information of the feature map. The combination of the two can reduce the computational complexity while maintaining the richness of features. The separated convolutional layer and pooling layer can be flexibly combined into different network structures, and the model architecture can be optimized by trying various configurations. Stronger adaptability: For specific tasks and datasets, separately designing the convolution and pooling strategies can improve the adaptability and performance of the model.

[0083] It can be understood that by automatically generating a network model through an algorithm, the workload of manual design can be significantly reduced, especially in scenarios where the network structure is complex or the requirements are changeable. The automated algorithm can quickly search a large number of possible architecture combinations and select the model structure with the optimal performance, thereby improving the model performance. Using the identification matrix and the information flow matrix, the network structure is dynamically adjusted according to the characteristics of the input data, so that the model can adapt to various task requirements. The automatic construction method can explore a wider network architecture design space, generate different levels and types of network combinations to meet complex application scenarios. The automated construction process is convenient for standardized operations and can be repeatedly applied to different tasks and datasets, significantly improving the development efficiency.

[0084] In this embodiment, the neural network structure is converted into a tree data structure, and the depth-first search algorithm is used to optimize the tree data structure to generate an executable neural network model, including: By traversing the levels of the neural network structure, the nodes of the neural network structure are defined as the nodes in the tree data structure, and the node is represented as a triple , the root node is defined as the node without any input operation, the leaf node is defined as the terminal node without child nodes in the tree data structure, and the value of the leaf node contains the set of operators and neural network nodes corresponding to the leaf node, and an initial tree data structure is constructed; where, is the value of the node, representing the state of the neuron; is the index of the node; is the set of child nodes; Starting from the root node, recursively traverse all child nodes and record the set of all paths from the root node to the leaf nodes :

[0085] where, is the node value in the path i in, is the length of the path; For the path set Add a dimensionality reduction layer and a linear layer to obtain an executable neural network model.

[0086] Among them, the recursive call of DFS is described as:

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

[0088] Add a dimensionality reduction layer and a linear layer to the optimized tree-shaped data structure to obtain an executable neural network model.

[0089] As Figure 3 and Figure 4 shown, Figure 3 This is the schematic diagram of the tree construction process of the operator and neural network node in the embodiment of the present invention. Figure 4 This is the schematic diagram of the process of executing the depth-first search algorithm in the embodiment of the present invention. In Figure 3 , the root node of the tree (None, -1None, -1) represents the original input, and the initial value is an empty operator. The nodes on each branch represent different operators (such as e, a, b, c, d) and their associated neural network nodes (such as 1, 2, 3). The operators are sequentially added to the tree according to the connection relationship in the identification matrix, forming a complete path from the root node to the leaf node. In Figure 4 , the process of traversing the tree structure through the depth-first search (DFS) algorithm: the starting point is the root node of the tree (e, 0), indicating that the data starts to be transmitted from the input node. The algorithm visits all branches along the depth of the tree, preferentially delving into each path until reaching the leaf node (such as c, 3). The process of the path from the root node to the leaf node records the complete connection relationship between the neural network nodes and the operators, and finally forms a directly trainable network model.

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

[0091] In the formula, , is the true value, is the predicted value of the model; if there are samples, the overall expression of the loss function is:

[0092] Update the training parameters of the executable neural network model using the Adam optimizer, including: Calculate the first moment estimate at the current t moment according to the following formula , representing the exponentially weighted average of the gradients:

[0093] where is the exponential decay rate of the first moment, set to 0.9, is the current gradient, is t- the first moment estimate at time 1; Calculate the second moment estimate at the current t moment according to the following formula , representing the exponentially weighted average of the squared gradients:

[0094] where is the exponential decay rate of the second moment, set to 0.9; Perform bias correction on the first moment estimate and the second moment estimate according to the following formula:

[0095]

[0096] where is the bias correction value of the first moment, is the bias correction value of the second moment, is to the power of t , is to the power of t ; Update the training parameters of the executable neural network model according to the following formula:

[0097] where is the value of the training parameter after the t th iteration, is the value of the training parameter at the th iteration, is the learning rate, is a constant to prevent division by zero, set to ; Use a cosine annealing learning rate scheduler to update the learning rate of the training of the executable neural network model, and its expression is as follows:

[0098] Among them, is the learning rate of cosine annealing, indicating the learning rate used in the t th training, is the minimum learning rate, is the initial learning rate, is the training period, is the current training round.

[0099] The mean squared error loss function is selected as the loss function for the test set to measure the accuracy of the model, and its definition is as follows:

[0100] In the formula, is the true value, is the predicted value of the model, is the number of samples.

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

[0102] It can be understood that the cosine function scheduler is used to dynamically adjust the learning rate, gradually decreasing the learning rate as the training progresses, making the optimization process more refined. By periodically restarting the learning rate, exploration is carried out near the local optimum to help the model quickly find the global optimum. The learning rate restart mechanism can prevent the model from stopping training due to too low a learning rate, ensuring a more sufficient optimization process. The scheduler supports multiple restarts, enabling the model to be optimized with different strategies at different training stages, thus adapting to complex loss surfaces.

[0103] As Figure 5 shown, Figure 5 is a schematic diagram of the change of the loss function during the training process of the embodiment of the present invention, Figure 5 showing the changes of the training loss and the validation loss with the number of rounds during the training process of the model. The training loss (blue curve) drops rapidly in the initial stage, indicating that the fitting effect of the model on the training set is gradually improved, and then tends to be stable and remains at a low level, showing good convergence of the model. The validation loss (orange curve) has relatively small fluctuations as a whole. Although there are certain fluctuations, it generally remains stable and within a low range, reflecting that the model has good generalization ability and robustness. The difference between the training loss and the validation loss is small, indicating that the model does not show overfitting. Overall, with the settings of the optimizer (such as Adam) and the learning rate scheduler (such as cosine annealing), the model can effectively learn features and optimize performance. In this embodiment, after the executable neural network model is trained and before the trained neural network model is optimized using the memetic algorithm, the method further includes: According to the following formula, calculate the Pearson coefficient of the trained neural network model as the initial fitness value of the memetic algorithm:

[0104] Where: is the number of samples, and the variable value is the predicted value of the neural network model, and the variable value is the true breeding phenotype label value. is and the sum of the products of the corresponding values. is the sum of the variable values ; is the sum of the variable values ; is the sum of the squares of the variable values ; is the sum of the squares of the variable values ;

[0105] In this embodiment, as shown in Figure 6 , Figure 6 is the flow chart of the memetic algorithm executed in the embodiment of the present invention. Use the memetic algorithm to optimize the trained neural network model to obtain the optimal breeding phenotype prediction model. The specific steps include: Step a: Use two trained neural network models as individuals in the population, and select individuals from the population using the roulette wheel selection algorithm according to the initial fitness value.

[0106] Among them, selecting individuals from the population using the roulette wheel selection algorithm according to the initial fitness value includes: According to the following formula, calculate the sum of the fitness values of all individuals :

[0107] Where: represents the th individual, represents calculating the fitness value of; According to the following formula, calculate the probability that the individual is selected:

[0108] According to the following formula, construct the cumulative probability distribution :

[0109] In the formula, represents the cumulative probability of the selected individual , generate a random number , and obtain the individual that satisfies as the selected individual.

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

[0111] As an exemplary illustration, take the two selected individuals as the parental individuals and parental individual . Select the crossover point and alternately exchange the gene segments of parental individual and parental individual to generate two offspring individuals as new candidate individuals, and their expressions are as follows:

[0112]

[0113] wherein, and are the offspring solutions generated after crossover from parental individual and .

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

[0115] As an exemplary illustration, performing a mutation operation on the new candidate individuals includes two methods: The first is to randomly select a gene locus in the new candidate individual and change the value of this gene locus to another value.

[0116] As an exemplary illustration, assume that the genes of individual are represented as , and perform mutation at locus . The mutated individual is:

[0117] wherein, is a newly generated random value.

[0118] The second is to randomly select two gene loci between the two new candidate individuals and exchange them.

[0119] As an exemplary illustration, randomly select two gene loci and , swap their values. If the individual , the mutated individual is:

[0120]

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

[0122] Step d: Perform local search on the individuals after the mutation operation, and recalculate the fitness value of the individuals.

[0123] For each neural network model in the individuals after the mutation operation, perform a separation operation to obtain a convolution information flow containing only convolution operators and a pooling information flow containing only pooling operators.

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

[0125] In the formula, is the set of convolutional layers, is the set of pooling layers.

[0126] Control the number of iterations to itro, and loop to perform the mutation operation in this embodiment on each pooling information flow to obtain a new pooling information flow. Merge the separated convolution information flow and the new pooling information flow, and replace the current neural network model with the new neural network model. Among them, the merge operation between information flows is defined as follows:

[0127] Among them, error means throwing an error, represents the separated convolution information flow, represents the mutated pooling information flow, is the neural network model after merging the separated convolution information flow and the mutated pooling information flow.

[0128] Use the new neural network model to form a new individual, and calculate the fitness value of the new individual.

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

[0130] As an exemplary illustration, the elite selection process is as follows: Construct the set , where i is the index of the individual.

[0131] According to the fitness score Sort the set S in descending order of fitness score to obtain a new set .

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

[0133] According to the following formula, split the elite individual set for splitting:

[0134] where is the current population, is the fitness score, is the model set, is the node matrix set. Define the function , where is the set of individuals selected through elite selection.

[0135] As an exemplary illustration, the tournament selection process is as follows: Randomly select tournament individuals from the current population P according to the following formula:

[0136] In the formula, is a random function, is the index of the individual selected in the th tournament.

[0137] Let , where is the subset of individuals in the th tournament.

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

[0139] Form a set from the set of winners of each tournament to form a new population.

[0140] where is the current population, is the fitness score, is the model set, is the node matrix set. Define the random selection function , where is the set of indices of the selected individuals.

[0141] It can be understood that the memetic algorithm balances efficiency and accuracy in global and local searches by adjusting the search strategy according to the objective function. The memetic algorithm emphasizes maintaining the diversity of the population to avoid premature convergence, thus exploring the search space more widely. The memetic algorithm can jump out of local optima and find better global solutions through diverse population search strategies. Based on global search, the accuracy and generalization ability of the solution are further improved through local mutation and optimization. When optimizing complex network structures, the memetic algorithm can more effectively find model parameters with higher fitness, thereby enhancing the prediction performance of the overall model.

[0142] Table 1 Comparison of prediction accuracies of breeding phenotypes of different models

[0143] As shown in Table 1, Table 1 presents a comparison of the prediction accuracies of different models in two environments, namely TKW (plant height) and GP (grain hardness). Among them, the MA model of the present invention shows the best performance, achieving Pearson correlation coefficients of 0.69 and 0.52 respectively, indicating its best comprehensive performance in predicting complex traits. In contrast, the traditional method GBLUP has the lowest accuracy, with correlation coefficients of 0.18 and 0.12 in the TKW and GP environments respectively, indicating that it is difficult to effectively process large-scale gene data. The traditional machine learning models LightGBM and SVR show medium performance. Although LightGBM reaches 0.64 in the TKW environment, it fails to surpass the MA model. Deep learning-based models such as DeepGS and DNNGP perform well in both environments. Among them, the accuracy of DNNGP is 0.67 and 0.5 respectively, close to the performance of the MA model but still slightly inferior. DLGWAS shows relatively stable performance (0.63 and 0.48) in both environments but also fails to surpass the MA model. Generally speaking, the MA model combines global search and local optimization capabilities, demonstrating significant advantages in prediction performance and verifying its application value and reliability in modern breeding research.

[0144] It is understandable that the present invention proposes a new breeding phenotype prediction model. By introducing a multi-level and complex structure, it aims to effectively address various deficiencies in traditional breeding phenotype prediction methods. The model adopts a deep learning architecture, capable of capturing the non-linear relationship between traits and genotypes, overcoming the limitation of linear mixed models that cannot 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) has been significantly improved. At the same time, the model utilizes multi-channel input and a flexible structure design, showing stronger robustness to changes in input data and being able to handle diverse biological data. In addition, the depth and complexity of the model structure endow it with good scalability, suitable for different scenarios, and exceeding the limitations of simple models in specific tasks. This comprehensive approach not only improves the prediction performance but also provides a better biological interpretation of the relationship between genes and traits, demonstrating important application value in modern breeding and genomics research.

[0145] The present invention automatically extracts important features from high-dimensional and complex data through multi-layer convolution and deep learning mechanisms, and combines the MA intelligent algorithm to optimize the model structure, significantly improving the efficiency and accuracy of feature learning. At the same time, the model uses non-linear activation functions to effectively capture the complex non-linear relationship between genotypes and traits, and ensures the consistency of data shape through the forward propagation mechanism, thereby enhancing the robustness to input data. In the prediction of complex traits (such as feed conversion ratio, yield, disease resistance, etc.), the model combines the DFS depth-first algorithm and comprehensive consideration of multiple genes and environmental factors, showing higher prediction accuracy. In addition, the model has strong flexibility and adaptability. Through the decoupled connection and stacking of operators and neural network nodes, it can be adjusted and optimized according to task requirements, and is applicable to 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 powerful technical support for precision agriculture and personalized breeding.

[0146] Example 3 As Figure 7 shown, this example 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 example, and includes: an acquisition module 100, a division module 200, a construction module 300, an allocation module 400, a conversion and optimization module 500, a training and optimization module 600, and a prediction module 700.

[0147] Among them, the acquisition module 100 is used to acquire genotype data and phenotype data, and construct a data set by using the genotype data and the phenotype data; the division module 200 divides the data set into a training set and a test set; the construction module 300 is used to construct a convolutional operator space, a pooling operator space, and neural network nodes; the allocation module 400 is used to allocate convolutional operators and pooling operators for the neural network nodes from the convolutional operator space and the pooling operator space to generate a neural network structure; the conversion and optimization module 500 is used to convert the neural network structure into a tree data structure, and optimize the tree data structure by using a depth-first search algorithm to generate an executable neural network model; the training and optimization module 600 is used to train the executable neural network model by using the training set, and optimize the trained neural network model by using a memetic algorithm to obtain an optimal breeding phenotype prediction model; the prediction module 700 is used to input the test set data into the optimal breeding phenotype prediction model for processing to obtain a breeding phenotype prediction result.

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

[0149] In the description of this specification, the descriptions with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or N embodiments or examples in a suitable manner. In addition, without contradiction, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples.

[0150] In addition, the terms "first" and "second" are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "N" is at least two, such as two, three, etc., unless otherwise specifically defined.

[0151] Any process or method description depicted in a flowchart or otherwise described herein can be understood to represent a module, segment, or portion of code including one or more executable instructions for implementing a customized logical function or process. The scope of the preferred embodiments of the present invention includes additional implementations where functions may be executed not in the order shown or discussed, including in a substantially simultaneous manner or in the reverse order according to the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.

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

[0153] Those of ordinary skill in the art of the present technology can understand that all or part of the steps carried by the methods of the above embodiments can be completed by instructing relevant hardware through a program. The program can be stored in a computer-readable storage medium. When the program is executed, it 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 embodiments of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the embodiments here. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present invention should be included within the protection scope 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 data and phenotype data, and construct a data set using the genotype data and the phenotype data; Divide the dataset into training and testing sets; Construct convolution operator space, pooling operator space and neural network nodes; Allocating 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; Converting the neural network structure into a tree data structure, and optimizing the tree data structure using a depth-first search algorithm to generate an executable neural network model; Using the training set, the executable neural network model is trained, and the trained neural network model is optimized using a meme algorithm to obtain an optimal breeding phenotype prediction model; The test set data is input into the optimal breeding phenotype prediction model for processing to obtain breeding phenotype prediction results.

2. The breeding phenotype prediction method for gene sequence data according to claim 1, characterized in that: After dividing the data set into a training set and a test set, the method further includes: For the even-dimensional feature data in the training set, it is encoded according to the following formula: For the odd-dimensional feature data in the training set, it is encoded according to the following formula: in, Represents the location index of the sample data. represents the dimension index of the feature vector, Represents the dimension of the feature vector.

3. The breeding phenotype prediction method for gene sequence data according to claim 1, characterized in that: According to the following formula, the convolution operator space and pooling operator space are constructed: in, Represents a character set, each character in the set is used to identify a convolution operator or pooling operator; is a set of convolution operators, characters They are different convolution operators, characters Indicates no connection or no operation; is a pooling operator set, character They are different pooling operators. To store a set of operators selected from the convolution operator space or the pooling operator space, is the convolution operator space, is the pooling operator space, Indicates the use of character mapping to the corresponding convolution operator or pooling.

4. The breeding phenotype prediction method for gene sequence data according to claim 1, characterized in that: The neural network node includes a CBAN attention mechanism and a plurality of residual connected convolution blocks; The CBAN attention mechanism includes a channel attention mechanism and a spatial attention mechanism; The channel attention mechanism is used to weight the weights of different channels of input data, including: According to the following formula, the channel attention weight is calculated : in, represents the feature map of the input data, is the number of channels, and is the height and width of the feature map; and are different learning weight matrices, is the reduction rate, is the activation function, Sigmoid activation function; Generate the channel attention weight feature map according to the following formula : in, Represents element-wise dot product; The spatial attention mechanism is used to weight the spatial position of the input data, including: According to the following formula, the spatial attention weight is calculated : in, Represents the feature map Perform average pooling in the channel dimension, Indicates the maximum pooling of the feature map in the channel dimension, and concat() means connecting two feature maps in 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 neural network nodes are connected to the original input or The connection method of neural network nodes is the index position information of the neural network node, which is used to determine the Neural network nodes and Is there a connection between them? Represents the computational operation unit of the current neural network node.

5. The breeding phenotype prediction method for gene sequence data according to claim 1, characterized in that: 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 the nodes are represented as triples. , define the root node as a node without any input operation, and the leaf node as a terminal node without child nodes in the tree data structure, and the value of the leaf node contains the set of operators and neural network nodes corresponding to the leaf node, and construct the initial tree data structure; where, is the value of the node, indicating the state of the neuron; is the index of the node; is a collection of child nodes; Starting from the root node, recursively traverse all child nodes and record all the path sets from the root node to the leaf nodes. : in, For path i The node values ​​in is the length of the path; For path collection Add dimensionality reduction layers and linear layers to get an executable neural network model.

6. The breeding phenotype prediction method for gene sequence data according to claim 1, characterized in that: Constructing a loss function and using an optimizer and a learning rate scheduler to train the executable neural network model, including: The smooth L1 loss function is used as the loss function for training, and its expression is as follows: In the formula, , is the true value, is the predicted value of the model; Use the Adam optimizer to update the training parameters of the executable neural network model, including: According to the following formula, calculate the current t First moment estimate of time , which represents the exponentially weighted average of the gradient: in, is the exponentially weighted decay rate of the first-order moment, is the current gradient, for t- First moment estimate at time 1; According to the following formula, calculate the current t Second moment estimate of time , which represents the exponentially weighted average of the squared gradient: in, is the exponentially weighted decay rate of the second-order moment; According to the following formula, the first-order moment is estimated and the estimate of the second moment To perform bias correction: in, is the deviation correction value of the first-order moment, is the bias correction value of the second-order moment, for of t Power, for of t Power; The training parameters of the executable neural network model are updated according to the following formula: in, For the t The training parameter value after iterations, For the The training parameter value at the iteration, is the learning rate, A constant to prevent division by zero; Use the cosine annealing learning rate scheduler to update the learning rate of the executable neural network model training. The expression is as follows: in, is the learning rate of cosine annealing, indicating that t The learning rate used in the training, is the minimum learning rate, is the initial learning rate, is the training cycle, is the current training round.

7. The breeding phenotype prediction method for gene sequence data according to claim 1, characterized in that: After completing the training of the executable neural network model and before optimizing the trained neural network model using the meme algorithm, the method further includes: According to the following formula, the Pearson coefficient of the trained neural network model is calculated as the initial fitness value of the meme algorithm: in, is the number of samples, the variable value is the predicted value of the neural network model, the variable value is the true breeding phenotype tag value, for and The sum of the products of the corresponding values, For variable value The sum of For variable value The sum of For variable value The sum of the squares of For variable value The sum of squares.

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

9. The breeding phenotype prediction method based on gene sequence data according to claim 8, characterized in that: Based on the initial fitness value, individuals are selected from the population using the roulette wheel selection algorithm, including: According to the following formula, calculate the sum of the fitness values ​​of all individuals : in, Indicates Individuals, Representation calculation The fitness value of According to the following formula, calculate the individual Probability of being selected : According to the following formula, the cumulative probability distribution is constructed : In the formula, Indicates the selection of individuals The cumulative probability of Generate random numbers , get satisfaction Individual As a chosen individual; Perform crossover operations on the selected individuals to generate new candidate individuals, including: The two selected individuals are used as parent individuals and parent individuals ; Select the crossover point and alternately exchange the parent individuals and parent individuals Gene segments of the gene are generated, and two offspring individuals are generated as new candidate individuals; Perform mutation operations on new candidate individuals, including: Randomly select a gene locus in a new candidate individual and change the value of the gene locus to another value, or randomly select two gene loci between two new candidate individuals for exchange; Perform local search on the individuals after mutation operation and recalculate the fitness value of the individuals, including: For each neural network model in the individual after the mutation operation, a separation operation is performed to obtain a convolution information flow containing only convolution operators and a pooling information flow containing only pooling operators; Perform iterative mutation operation on the pooled information flow, and merge the mutated pooled information flow and the convolution information flow to obtain a new neural network model; Use the new neural network model to form new individuals and calculate the fitness values ​​of the new individuals.

10. A breeding phenotype prediction system for gene sequence data, characterized in that: include: An acquisition module, used to acquire genotype data and phenotype data, and construct a data set using the genotype data and the phenotype data; The partitioning module divides the data set into training set and test set; Construction module, used to construct convolution operator space, pooling operator space and neural network nodes; An allocation module, 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; A conversion optimization module, 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; A training optimization module, used to train the executable neural network model using a training set, and optimize the trained neural network model using a meme algorithm to obtain an 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 to obtain the breeding phenotype prediction result.

Images