Breeding method based on convolution-long and short term memory neural network

By adjusting the structure of the convolutional-long short-term memory neural network and using appropriate evaluation criteria, the problem of low prediction accuracy in genomic selection technology was solved, resulting in a more efficient breeding method that adapts to the characteristics of genomic data and genetic laws, thereby improving breeding efficiency.

CN119028449APending Publication Date: 2024-11-26SHENYANG AGRI UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411106976.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-13
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

Existing whole-genome selection technology has low prediction accuracy when dealing with complex interactions between genes and a large number of low-efficiency gene markers, and traditional deep learning models are not suitable for evaluation criteria in breeding, which affects breeding efficiency.

Method used

A breeding method based on convolutional long short-term memory neural networks was adopted. By adjusting the network structure according to the number of gene markers, combined with median filtering for noise reduction and using the MSE and Pearson correlation coefficient evaluation model, a deep learning breeding method adapted to the characteristics of genomic data was constructed.

Benefits of technology

It improves breeding efficiency, enhances prediction accuracy and model interpretability, conforms to genetic principles, adapts to different genomic datasets, and is suitable for breeding decision-making.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119028449A_ABST
    Figure CN119028449A_ABST
Patent Text Reader

Abstract

The invention provides a breeding method based on a convolution-long and short term memory neural network, and the method comprises the steps: carrying out the phenotype and genotype processing of a target material, and obtaining a training data set, which comprises genotype data and phenotype data; inputting the training data set into the convolution-long and short-term memory neural network for training and verification to obtain a trained prediction model; inputting to-be-detected data into the prediction model to obtain a breeding value (or a phenotype value) of a variety with an unknown phenotype; and guiding a breeding decision by using a prediction result. According to the characteristics that genome data contains noise, has a position effect and can generate large difference due to species and a sequencing method, a convolution-long and short-term memory neural network framework of which the structure is changed according to the number of gene markers and a corresponding noise filtering and model training result evaluation method are provided; the deep learning breeding method which has better interpretability, conforms to genetic laws and has universality is constructed, and the breeding efficiency is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a breeding method based on a convolution-long short-term memory neural network. Background Art

[0002] As a new breeding method, whole genome selection (GS) technology uses mathematical models to evaluate the relationship between all given genotype molecular markers and phenotypes to infer the estimated breeding value (BLUP value) of the genome, and use the inference results to predict individuals with unknown phenotypes. This method can shorten the breeding cycle and eliminate invalid parent combinations in advance without clarifying the genetic mechanism in advance. Since all molecular markers are evaluated, it is more efficient and fast to aggregate the minor effect genes that control quantitative traits. At present, GS has been widely used to analyze phenotypes through single nucleotide polymorphism sites (SNPs) using algorithms such as genome best linear unbiased estimate (GBLUP), ridge regression best linear unbiased estimate (rrBLUP), and Bayesian. However, experiments have shown that a large number of minor effect gene markers will reduce the prediction accuracy of GS, and the nonlinear relationship between genes and phenotypes also hinders the improvement of the prediction accuracy of the GS algorithm. Breeders need more effective tools for data processing and analysis.

[0003] Machine learning is currently a popular field of artificial intelligence research. Its purpose is to automatically generate algorithms based on input data to solve tasks. Machine learning has achieved breakthrough results in tasks such as image recognition, speech analysis, and text classification. In 2006, Geoffrey Hinton proposed deep learning based on machine learning, which provides the model with stronger learning capabilities. Unlike traditional whole genome selection models, the strong learning ability of machine learning models can better perform data mining, that is, to summarize and extract hidden information in the data without assuming the prior distribution of gene marker effects. The nonlinear activation function used in deep learning can also better handle nonlinear problems. Therefore, machine learning, especially deep learning, is the most suitable tool for analyzing genomic data that hides a large amount of information about gene interactions and has an increasing number of sites.

[0004] At present, the deep learning model that is more mature in GS is the convolutional neural network. These models usually extract hidden information from genomic data by a convolution layer, and then integrate the extracted information through a fully connected layer to obtain phenotypic values. The convolution layer is composed of several convolution kernels. The size and moving step of the convolution kernel determine the number of features extracted by the convolution kernel at one time and the distance between the extracted feature fragments respectively. The number of convolution kernels determines how many aspects a convolution layer summarizes features from. Therefore, after passing through the convolution layer, low-level features become high-level features with mutual relationships, and the closer the distance, the stronger the relationship. After multiple convolutions, a high-level feature represents multiple low-level features similar to it, and the result is obtained through the fully connected layer. However, there are complex interactions between genes. Two interacting genes may be separated by tens of thousands of bases or located on different chromosomes, which obviously exceeds the size of the convolution kernel. The number of genetic markers will vary greatly depending on the population, sequencing method, and species, so a reasonable structure is also needed to adapt to different data sets and make it more universal. In addition, the loss function MSE used in deep learning can make the model predict values ​​​​fit to the true values ​​during the iteration process, but the Pearson correlation coefficient is usually used as an evaluation criterion in breeding, so a result evaluation method that is more suitable for the breeding direction is needed. Summary of the invention

[0005] In order to overcome the deficiencies of the prior art, the object of the present invention is to provide a breeding method based on a convolutional-long short-term memory neural network that changes structure according to the number of gene markers, as well as a data denoising and model evaluation method.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A breeding method based on convolution-long short-term memory neural network, comprising:

[0008] Obtaining a target material, and performing phenotype and genotype determination and processing on the target material to obtain a training data set; the training data set includes genotype data and phenotype data;

[0009] The training data set is input into a convolutional-long short-term memory neural network whose structure is changed according to the number of genes for training, and the MSE and Pearson correlation coefficient are used for simultaneous testing to obtain a trained prediction model;

[0010] Inputting the data to be tested into the prediction model to obtain a prediction result; the prediction result is used to predict the phenotype of an individual with an unknown phenotype;

[0011] The prediction results are used to guide breeding decisions.

[0012] Preferably, obtaining a target material and performing phenotype and genotype determination on the target material to obtain a training data set includes:

[0013] Obtaining the phenotype and SNP sequence of the target material;

[0014] The BLUP values ​​of the phenotypic measurements of the target materials at different locations and years were calculated using the META-R software to obtain the phenotypic data.

[0015] Using genotyping to obtain SNP sites in the gene sequence, and screening the SNP sites according to allele frequency, deletion rate and significance, using Beagle to fill in the deletions in the remaining sites, and digitizing the filled SNP sites to obtain the genotype data;

[0016] Preferably, after obtaining the target material and measuring the phenotype and genotype of the target material to obtain the training data set, the method further includes:

[0017] Using a filter window to detect noise points on the gene sequence of the training data set to obtain a noise mean;

[0018] Calculate the mean of the gene sequence in the filtering window and replace the middle gene in the window with the mean, and perform denoising on the gene sequence of the training data set in the corresponding filtering window;

[0019] Slide the filter window and return to the step of "calculating the mean of the gene sequence in the filter window and replacing the middle gene in the window with it, and denoising the gene sequence of the training data set in the corresponding filter window" until the entire gene sequence of the training data set is traversed to obtain the denoised genotype data.

[0020] Preferably, before the training data set is input into the convolutional-long short-term memory neural network for training and verification, the model hyperparameters of the convolutional-long short-term memory neural network are initialized; the model hyperparameters include: learning rate, batch size and dropout rate.

[0021] Preferably, the convolutional-long short-term memory neural network includes a convolutional layer, a long short-term memory network and a fully connected layer connected in sequence.

[0022] Preferably, the convolution-LSTM neural network automatically determines the network structure according to the number of SNP markers. When there are many markers, four convolution kernels with a size of one thousandth of the number of SNP sites and a step length of one ten-thousandth of the number of SNP sites are used for preliminary convolution, while when there are fewer markers, the size of the convolution kernel is one thousandth of the number of SNP sites, and the step length is fixed to 1, followed by a normalization layer to standardize the output data and a Relu activation function; then input a layer of bidirectional LSTM neural network; the output result is flattened and input into the fully connected layer to obtain the final predicted phenotype.

[0023] Preferably, the loss function of the convolution-long short-term memory neural network is the mean square error between the predicted phenotype and the true value, and the Pearson correlation coefficient is used to evaluate the model from the breeding requirements.

[0024] Preferably, the data to be tested is input into the prediction model to obtain a prediction result; the prediction result is used to predict the phenotype of an individual with an unknown phenotype;

[0025] Preferably, breeding decisions are guided according to the prediction results, varieties with good phenotypes are planted, and inferior varieties are eliminated.

[0026] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0027] The present invention provides a breeding method based on a convolution-long short-term memory neural network, comprising: obtaining a target material, and measuring and processing the phenotype and genotype of the target material to obtain a training data set, wherein the training data set includes genotype data and phenotype data; inputting the training data set into a convolution-long short-term memory neural network for training, and using MSE and Pearson correlation coefficient to verify from two perspectives of deep learning and breeding fields to obtain a trained prediction model; inputting the data to be tested into the prediction model to obtain a phenotypic prediction result of an unknown phenotype variety; and using the prediction result to guide breeding decisions. According to the characteristics that genomic data contains noise, has a position effect, and has large differences due to species and sequencing methods, the present invention proposes a convolution-long short-term memory neural network framework that changes the structure according to the number of gene markers, and a corresponding noise filtering and model training result evaluation method, thereby constructing a deep learning breeding method with better interpretability, conforming to genetic laws, and universality, thereby improving breeding efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0029] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0030] Figure 2 A schematic diagram of a model structure provided by an embodiment of the present invention;

[0031] Figure 3 A schematic diagram of a technical route provided for an embodiment of the present invention;

[0032] Figure 4A schematic diagram of the changing trend of MSE and Pearson correlation coefficient during the training process provided by an embodiment of the present invention (one repetition). DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] The purpose of the present invention is to provide a breeding method based on a convolution-long short-term memory neural network that is adapted to genome data and a data denoising and model evaluation method.

[0035] Based on the characteristics that genomic data contains noise, has position effects and can produce large differences due to species and sequencing methods, a convolutional-long short-term memory neural network framework that changes its structure according to the number of genetic markers is proposed, as well as a corresponding method of using median filtering to filter noise and using MSE and Pearson correlation coefficient to evaluate model training results. In this way, a deep learning breeding method with better interpretability, conforming to genetic laws and universality is constructed, which improves breeding efficiency.

[0036] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] Figure 1 A flow chart of a method provided by an embodiment of the present invention, such as Figure 1 As shown, the present invention provides a breeding method based on convolution-long short-term memory neural network, comprising:

[0038] Step 100: Obtain a target material, and perform phenotype and genotype determination and processing on the target material to obtain a training data set; the training data set includes genotype data and phenotype data;

[0039] Step 200: input the training data set into a convolutional-long short-term memory neural network that automatically changes its structure according to the number of genes for training, and use MSE and Pearson correlation coefficient to simultaneously test to obtain a trained prediction model;

[0040] Step 300: input the data to be tested into the prediction model to obtain a prediction result; the prediction result is used to predict individuals with unknown phenotypes;

[0041] Step 400: Using the prediction results to guide breeding decisions.

[0042] Preferably, obtaining a target material and performing phenotype and genotype processing on the target material to obtain a training data set includes:

[0043] The BLUP values ​​of the phenotypic measurements of the target materials at different locations and years were calculated using the META-R software to obtain the phenotypic data.

[0044] Using genotyping to obtain SNP sites in the gene sequence, and screening the SNP sites according to allele frequency, deletion rate and significance, using Beagle to fill in the deletions in the remaining sites, and digitizing the filled SNP sites to obtain the genotype data;

[0045] Preferably, after obtaining the target material and measuring the phenotype and genotype of the target material to obtain the training data set, the method further includes:

[0046] De-noising the data in the training data set to obtain a denoised training data set;

[0047] Preferably, denoising the data in the training data set to obtain a denoised training data set includes:

[0048] Using a filter window to detect noise points on the gene sequence of the training data set to obtain a noise mean;

[0049] Calculate the mean of the gene sequence in the filtering window and use it to replace the middle gene in the window, and denoise the gene sequence of the training data set in the corresponding filtering window;

[0050] Slide the filter window and return to the step of "calculating the mean of the gene sequence in the filter window and replacing the middle gene in the window with it, and denoising the gene sequence of the training data set data in the corresponding filter window" until the entire training data set gene sequence is traversed to obtain the denoised training data set.

[0051] Preferably, the training data set is input into a convolutional-long short-term memory neural network whose structure is changed according to the number of genes for training, and the MSE and Pearson correlation coefficient are used to simultaneously test to obtain a trained prediction model;

[0052] Preferably, the data to be tested is input into the prediction model to obtain a prediction result; the prediction result is used to predict the phenotype of an individual with an unknown phenotype;

[0053] Preferably, breeding decisions are guided according to the prediction results, varieties with good phenotypes are planted, and inferior varieties are eliminated.

[0054] Specifically, the specific contents in this embodiment include:

[0055] 1) Acquisition and processing of phenotype and genotype

[0056] Phenotype is the BLUP value of the phenotypic measurement value of the target material at different locations and years.

[0057] The SNP sites obtained by genotyping were used as genotype data. The genotypes were screened according to allele frequency, missing rate and significance. Beagle was used to fill in the missing sites in the remaining sites. The genotype data after filling were digitized, with the major allele assigned a value of 2, the heterozygous site assigned a value of 1, and the minor allele assigned a value of 0.

[0058] The phenotypes are retained to three significant digits to reduce noise.

[0059] Genotype sequences were denoised using median filtering.

[0060] This embodiment uses a filter window to filter the training data set, which can separate outliers in the training data set, making the data more accurate.

[0061] In this embodiment, Python data processing libraries such as Pandas and Numpy can be used to implement the denoising process. Pandas can easily process and operate sequence data, while Numpy can be used for efficient numerical calculations. The sliding window in the denoising process can be implemented by the rolling() function of Pandas, and the detection and processing of noise can be completed by custom functions.

[0062] 2) GS model training

[0063] The digitized phenotype and genotype are input into the model, the model parameters are set, and training begins.

[0064] Main model parameters:

[0065] Learning rate: The optimization step size of each iteration. If the prediction accuracy does not achieve a better value after multiple iterations during training, it will be reduced proportionally. In the embodiment, it is set to 15 times.

[0066] Batch size: The number of samples extracted at one time for model training. Small batches will increase the generalization ability of the model and improve prediction accuracy. Properly increasing the batch size can increase the training speed. In the embodiment, it is set to 20.

[0067] Dropout rate: During the training process, some neurons are allowed to stop working with a certain probability to enhance the generalization ability of the model and prevent overfitting. In the embodiment, it is set to 0.01.

[0068] Model structure:

[0069] This model consists of a convolutional layer, a long short-term memory network, and a fully connected layer. The input genotype data is convolved in the convolutional layer, and the size and step length of the convolution kernel change automatically according to the number of SNPs. The main purpose of this layer is to compress the number of markers for different genetic data, reduce the calculation time, and extract the correlation between adjacent SNP markers to produce a more stable data structure. Then, the sequence information in the data is extracted through the long short-term memory network. Figure 2 As shown, the expression of the network is:

[0070] f t =σ(W f x t +W f h t-1 +b f )

[0071] i t =σ(W i x t +W i h t-1 +b i )

[0072] g t =σ(W g x t +W g h t-1 +b g )

[0073] O t =σ(W o x t +W o h t-1 +b o )

[0074] S t =g t ·i t +S t-1 ·f t

[0075]

[0076] Among them, f t 、i t , g t , O t 、h t and S t are the states of the forget gate, input gate, input node, output gate, intermediate output and state storage unit of the t-th feature, σ is the sigmoid activation function, is the tanh activation function. t 、i t, g t , O t By h t-1 and the t-th feature, and at the same time the storage unit S of the t-1th feature t-1 Determine the state S of the tth feature storage unit t Finally, O t and S t Calculate the output h of the tth feature t , and affects the output of t+1 features. Thus, the model can make a sequential summary of the above. In order to better fit the interaction between genes, information is processed and integrated from both directions. Finally, the sequence information is input into the fully connected layer and the result is obtained. All samples are extracted once as an iteration, and the mean square error of the iteration is calculated as the evaluation criterion. When no better mean square error is obtained for 15 consecutive iterations, the learning rate is reduced. If no better mean square error is obtained for 60 consecutive iterations, the iteration output with the smallest mean square error is used as the model. By outputting the Pearson correlation coefficient on the training set during the iteration process, and using both numerical values ​​and trends as the basis for evaluating the model in the breeding field, it is more in line with breeding requirements.

[0077] 3) Prediction of unknown groups

[0078] The output model is called to predict the phenotype of the population to be tested, and the varieties with good phenotypes are selected according to the prediction results. In this embodiment, the prediction is for individuals with known phenotypes, and the predicted phenotype is compared with the actual phenotype to evaluate the practicality of the model.

[0079] like Figure 3 As shown, in the technical route of the present invention, the first bracket is step 1 phenotype and genotype processing, the second bracket is step 2 model training, and the third bracket is step 3 unknown population prediction and model evaluation. Exemplarily, this embodiment provides the following implementation:

[0080] 1) Phenotypic and genotypic treatment

[0081] The phenotypes and SNP sequences of the CUBIC population were downloaded from the Maize Group-Resources (maizego.org) website. This population is a double-row hybrid population. The data set provides phenotypic and genotypic data. Tassel was used to screen and retain SNP sites with a heterozygosity rate lower than 0.2 and a minimum allele frequency greater than 0.05. Two phenotypes, spike weight and plant height, were selected, and a mixed linear model was used for genome-wide association analysis to obtain the significance of SNP-phenotype correlation. About 9,000 SNPs with P ≤ 0.001 were selected for digitization for GS analysis.

[0082] The phenotypes are retained to three significant digits to reduce noise.

[0083] The significant SNPs were denoised using median filtering.

[0084] 2) GS model training

[0085] The numerical training set phenotype and genotype are input into the convolution-long short-term memory neural network. The network structure is: the network structure is determined according to the number of SNP markers. When there are many markers, four convolution kernels with a size of one thousandth of the number of SNP sites and a step length of one ten-thousandth of the number of SNP sites are used for preliminary convolution. When there are fewer markers, the size of the convolution kernel is one thousandth of the number of SNP sites, and the step length is fixed to 1. In this embodiment, the number of markers is large, so the first convolution layer is automatically used. The normalization layer is then connected to standardize the output data and a Relu activation function; then a 1-layer bidirectional long short-term memory neural network is input. The output result is flattened and input into the fully connected layer to obtain the final predicted phenotype. The mean square error between the predicted phenotype and the true value is calculated as the loss function of the neural network. The Adam optimizer is used. If a better mean square error is not obtained after 15 iterations, the function attenuation LambdaLR is used to reduce the learning rate. If no better results are obtained after 60 consecutive iterations, the optimal mean square error model is used as the result. The maximum number of iterations is 250. At the same time, the Pearson correlation coefficient on the training set during the iteration process is output to determine whether the model is suitable for subsequent breeding plans. In order to prevent overfitting, a dropout layer is added before and after the long short-term memory network. The initial learning rate, batch size and two dropout rates are 0.01, 20, 0.2, and 0.2, respectively. The average results of 5 10-fold cross-validation are used to evaluate the model performance, that is, 1 / 10 of the samples are extracted each time as unknown phenotype individuals for verification without duplication until all samples are used for verification. The cross-validation is repeated 5 times.

[0086] 3) Unknown population prediction and model evaluation

[0087] The obtained model is used to predict the genotype of the simulated unknown phenotype individual to obtain the predicted breeding value. In actual breeding, varieties with good phenotypes can be planted according to the predicted results, and inferior varieties can be eliminated. This embodiment calculates the Pearson correlation coefficient and mean square error between the predicted value and the true breeding value table to evaluate the practicality of the model. DNNGP, soyDNGP and GBLUP with the same parameters are used for comparison, as shown in Table 1.

[0088] Table 1

[0089]

[0090] The average prediction accuracy of plant height and ear grain weight predicted by convolution-long short-term memory neural network is 0.92 and 0.90, which is better than GBLUP, DNNGP and soyDNGP, with mean square errors of 267.06 and 75.24 respectively. It proves that recurrent neural network has good performance in predicting trends and true values ​​based on genomic data.

[0091] like Figure 4 As shown in the figure, MSE is not exactly the same as the Pearson correlation coefficient. By using the Pearson correlation coefficient to evaluate the training process, it can be intuitively reflected whether the model meets the breeding requirements. In this paper, the Pearson correlation coefficient of the model on the prediction set has reached stability and can be used to formulate subsequent breeding plans. If there is still an upward or downward trend, the model can be adjusted to fit or avoid overfitting by increasing or decreasing the number of overtraining. Using two evaluation indicators makes deep learning more suitable for application in the breeding field.

[0092] Whole genome selection has been looking for new models for breeding decisions, and the interpretability of the model is an important aspect. There are a large number of phenotype-independent genes in the genome, which act as noise to affect GS prediction. Therefore, the average effect of genes on phenotypes is averaged using mean filtering. CNN extracts local information through convolution kernels. From a biological perspective, it can be understood that convolution kernels extract local gene information. The high-level features obtained are the comprehensive effects of adjacent genes, and the high-level features are independent of each other. This information extraction method ignores long-distance gene interactions. Although multiple convolutional layers can be stacked to generalize to longer distances, long-distance gene interactions are a kind of hidden information that may be ignored in the initial convolution process. Recurrent neural networks focus on analyzing gene interactions. The output result is the comprehensive effect of the effect of the t-th gene and the upstream site effect. It can extract hidden sequence information to the maximum extent, and LSTM can flexibly determine the impact of the upstream site effect. Therefore, the model can better summarize the interaction between genes, and therefore achieve the best prediction accuracy and stable mean square error in prediction. At the same time, the MSE and Pearson correlation coefficient are used to evaluate the model at the same time, making the model more in line with breeding needs. MSE makes the predicted value closer to the true value, and the Pearson correlation coefficient makes the predicted value and the true value have the same trend. At the same time, the model training is completed from both numerical value and trend, which is more in line with breeding requirements.

[0093] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0094] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A breeding method based on convolutional-long short-term memory neural network, characterized by: Acquiring target materials, and performing phenotype and genotype determination and processing on the target materials to obtain a training data set; The training data set includes genotype data and phenotype data; The training data set is input into a convolutional-long short-term memory neural network whose structure is changed according to the number of genes for training, and the MSE and Pearson correlation coefficient are used for simultaneous testing to obtain a trained prediction model; Inputting the data to be tested into the prediction model to obtain a prediction result; the prediction result is used to predict the phenotype of an individual with an unknown phenotype; The prediction results are used to guide breeding decisions.

2. The breeding method based on convolution-long short-term memory neural network according to claim 1, characterized in that: Obtain target materials, and perform phenotype and genotype determination and processing on the target materials to obtain a training data set, including: Obtaining the phenotype and SNP sequence of the target material; the target material is a natural population, a biparental population or a multi-parental population, and the hybrid prediction uses a diallel hybridization or sparse diallel hybridization experimental design; Calculate the BLUP values ​​of the phenotypic measurements of the target materials at different locations and years by using META-R software to obtain the phenotypic data; The SNP sites of the gene sequence are obtained by genotyping, and the SNP sites are screened according to the allele frequency, deletion rate and significance. The deletions in the remaining sites are filled by Beagle, and the filled SNP sites are digitized to obtain the genotype data.

3. The breeding method based on convolution-long short-term memory neural network according to claim 1, characterized in that: After obtaining the target material, and measuring and processing the phenotype and genotype of the target material to obtain the training data set, the method further includes: Use a mean filter window to detect noise points on the gene sequence of the training data set to obtain a noise mean; Calculate the mean of the gene sequence in the filtering window and replace the middle gene in the window with the mean, and perform denoising on the gene sequence of the training data set in the corresponding filtering window; Slide the filter window and return to step "calculate the mean of the gene sequence in the filter window and replace the middle gene in the window with it, and denoise the gene sequence of the training data set in the corresponding filter window" until the entire gene sequence of the training data set is traversed to obtain the denoised training data set.

4. The breeding method based on convolution-long short-term memory neural network according to claim 1, characterized in that: Before inputting the training data set into the convolution-long short-term memory neural network for training and verification, it also includes: Initialize the model hyperparameters of the convolution-long short-term memory neural network; the model hyperparameters include: learning rate, batch size and drop rate.

5. The breeding method based on convolution-long short-term memory neural network according to claim 1, characterized in that: The convolution-long short-term memory neural network includes a convolutional layer, a long short-term memory network and a fully connected layer which are connected in sequence.

6. The breeding method based on convolution-long short-term memory neural network according to claim 2, characterized in that: The convolution-long short-term memory neural network is suitable for genetic data, and the generation of the convolution layer is automatically determined according to the number of SNP markers; when there are many markers, four convolution kernels with a size of one thousandth of the number of SNP sites and a step length of one ten-thousandth of the number of SNP sites are used for preliminary convolution; when there are fewer markers, the size of the convolution kernel is one thousandth of the number of SNP sites, the step length is fixed to 1, and the number is 4, followed by a normalization layer to standardize the output data and a Relu activation function; then a layer of bidirectional long short-term memory neural network is input; the output result is flattened and input into the fully connected layer to obtain the final predicted phenotype.

7. The breeding method based on convolution-long short-term memory neural network according to claim 1, characterized in that: The loss function of the convolutional-long short-term memory neural network is the mean square error between the predicted phenotype and the true value, and the Pearson correlation coefficient is used to verify the biological performance of the model.

Citation Information

Cited By

  • Large model data enhancement method and device

    CN120471109A