Rice phenotype prediction method and system based on whole genome selection

Abstract

The application belongs to the technical field of plant breeding, and particularly relates to a rice phenotype prediction method and system based on whole genome selection, which comprises the following steps: obtaining rice double-end sequencing data, performing genome alignment and variation detection to obtain SNP data, and performing data quality control; performing linkage disequilibrium analysis, population structure analysis and kinship analysis on the SNP data after quality control to screen SNPs; dividing the screened SNP set into a training set, a validation set and a test set in proportion, and training and validating the set using a convolutional neural network model based on an attention mechanism; introducing an attention mechanism module and a residual module, training five kinds of convolutional neural network models based on the attention mechanism of different data through five kinds of SNP set data, obtaining a prediction model trained by the best input data, and performing phenotype prediction; and the application can fully screen high-quality and representative SNP data, better capture genetic variation information related to a target phenotype, and improve the efficiency and prediction accuracy of the model.

Description

Technical Field

[0001] This invention belongs to the field of plant breeding technology, specifically relating to a method and system for predicting rice phenotypes based on whole-genome selection. Background Technology

[0002] With advancements in agricultural production demands and detection technologies, crop breeding has become a research hotspot. By analyzing and utilizing plant genomic information, we can more accurately select and improve individuals or varieties with superior traits, thereby achieving sustainable agricultural development. Genomic selection is a breeding method that utilizes high-density molecular markers covering the genome for selection. It can construct predictive models to predict and select individuals at an early stage to achieve the goal of breeding superior traits.

[0003] Traditional genomic selection methods utilize optimal linear unbiased estimation, treating individuals as random effects. They use a kinship matrix constructed from the genetic information of the training and prediction populations as the covariance matrix, estimating variance components through iterative methods, and then solving a mixture model to obtain the estimated breeding value for the individual to be predicted. However, a significant drawback of this method is that it ignores complex nonlinear relationships. Machine learning methods, on the other hand, can not only learn complex nonlinear relationships but also better capture latent patterns and features in genomic data. Various classic machine learning algorithms (such as SVR, XGBOOST, and RF) and custom convolutional neural networks have been widely applied to genomic selection.

[0004] The main drawbacks of current genomic selection methods include the following:

[0005] I. In genome selection, choosing the genomic data used to construct predictive models is a challenging issue. Selecting too few single nucleotide polymorphism (SNP) data may result in incomplete coverage of genetic variations, affecting the model's fitting ability and leading to poor predictive performance. On the other hand, selecting too many SNP data may introduce excessive noise—features that have no effect on the phenotype—which can also negatively impact the model's predictive performance, while significantly increasing computational resource requirements.

[0006] Second, current genomic selection methods do not take into account the importance of SNP features when constructing models, which may affect the accuracy of phenotypic prediction.

[0007] Third, existing deep learning prediction models may have certain limitations in the extraction and fusion of gene features, resulting in lower prediction accuracy. Summary of the Invention

[0008] In view of the above-mentioned problems, the present invention is proposed.

[0009] Therefore, the technical problem solved by this invention is: how to screen high-quality and representative SNP data, how to better capture genetic variation information related to the target phenotype, and improve the efficiency of the model and the accuracy of prediction.

[0010] To address the aforementioned technical problems, this invention provides the following technical solution: Acquire paired-end sequencing data of rice, perform genome alignment and variant detection to obtain SNP data, and conduct data quality control; utilize the quality-controlled SNP data to perform linkage disequilibrium analysis, population structure analysis, and phylogenetic analysis to screen out significant SNP sets; divide the screened SNP sets proportionally into training, validation, and test sets, and train and validate them using a convolutional neural network model based on an attention mechanism; introduce an attention mechanism module and a residual module, use the mean absolute error as the loss function, train using a stochastic gradient descent optimizer, perform a five-fold cross-validation on the model, calculate the mean of the coefficient of determination as the final performance index, and set an early stopping strategy to improve the model's generalization ability; train five different attention mechanism-based convolutional neural network models using five SNP sets, obtain the prediction model trained with the best input data, and perform phenotypic prediction.

[0011] As a preferred embodiment of the rice phenotypic prediction method based on whole-genome selection described in this invention, the following steps are included: performing genome alignment and variant detection, which involves comparing sequencing data with a rice reference genome to determine the position of each read on the genome, performing variant detection on the alignment results to obtain SNP data of rice; and performing quality control by checking the missing rate of rice samples after alignment with the obtained reference genome, and removing samples if the missing rate is greater than or equal to 0.2%.

[0012] For each SNP in the rice sample, we checked whether it lacked individual information. If the rate of missing individual information for a single SNP was greater than or equal to 0.2, it was directly screened out. We also checked the minor allele frequency for each SNP in the rice sample. If the minor allele frequency for a single SNP was less than or equal to 0.05, it was directly screened out. We screened out SNPs with high deletion rates and low minor allele frequencies, and finally filled the missing SNPs with major alleles.

[0013] As a preferred embodiment of the rice phenotypic prediction method based on whole-genome selection described in this invention, the linkage disequilibrium analysis includes performing linkage disequilibrium analysis by calculating and evaluating the relationship between linkage haplotype frequencies and SNP sites.

[0014] Among them, the chain disequilibrium value r 2 The calculation is expressed as:

[0015] D = P AB -P AP B

[0016]

[0017] Where A is the major allele of SNP1, B is the major allele of SNP2, and P AB P represents the frequency of AB as actually observed. A P B This represents the expected value of the AB frequency.

[0018] When r 2 A value of 1 indicates a complete linkage disequilibrium between the two SNPs; when r... 2 When the value is 0, it indicates that there is a complete linkage equilibrium between the two SNPs;

[0019] The population structure analysis includes using ADMIXTURE software to perform population structure analysis, setting key parameters such as population size and constructing a genetic correlation matrix. The population size is determined by the number of species contained in the sample.

[0020] The kinship analysis includes using a mixed linear model to perform association analysis on genotype and phenotype, using the genetic correlation matrix and kinship matrix of population structure as covariates to reduce potential confounding effects, and using different genome-wide association analysis thresholds (1e-1, 1e-2, 1e-3, 1e-4, 1e-5) to screen the set of SNPs.

[0021] By comparing the results of genome-wide association analysis with linkage disequilibrium, if a group of loci show significant association in the association analysis, and the linkage disequilibrium value r among the SNP loci is high, then the association is considered successful. 2 All values ​​are greater than or equal to the average attenuation distance, indicating a high degree of accuracy in the GWAS analysis results;

[0022] The average attenuation distance is expressed as:

[0023]

[0024] The average decay distance is the linkage disequilibrium value r among all SNP sites. 2 Half of the maximum value, where n is the number of pairwise combinations of SNPs.

[0025] As a preferred embodiment of the rice phenotypic prediction method based on whole-genome selection described in this invention, the step of training and validating the convolutional neural network model based on the attention mechanism includes using the selected SNP set as the input data of the model, dividing it into a training set, a validation set, and a test set in a ratio of 8:1:1, and training the convolutional neural network prediction model based on the attention mechanism.

[0026] An attention mechanism module is introduced, which weights the SNPs by multiplying the original input with the weights output by the attention module. The attention mechanism allows the model to dynamically learn and selectively focus on different SNPs.

[0027] The importance of each SNP to the prediction task is determined by calculating the attention weight of each SNP. The calculation formula is as follows:

[0028] Output=X*σ(W1(ReLU(W0(X)+b0))+b1)

[0029] The weights of input X are calculated through two fully connected layers, with parameters W0 and b0 and W1 and b1, respectively.

[0030] As a preferred embodiment of the rice phenotypic prediction method based on whole-genome selection described in this invention, the training and validation of the attention-based convolutional neural network model further includes introducing a residual module to obtain features at different scales of genotype data, using a convolutional layer with a kernel size of 20 to capture larger-scale gene features, using a convolutional layer with a kernel size of 4 to capture smaller-scale gene features, generating feature maps of different sizes, combining gene features at different scales as input to the next layer, and then stacking and fusing feature maps of different scales by channel.

[0031] As a preferred embodiment of the rice phenotypic prediction method based on whole-genome selection described in this invention, the training and validation of the attention-based convolutional neural network model further includes using the mean absolute error as the loss function, training with a stochastic gradient descent optimizer, setting the learning rate to 1×10⁻³, the weight decay coefficient to 1×10⁻⁶, the momentum to 0.90, the epoch to 200, performing a five-fold cross-validation on the model, and taking the mean of the five coefficients of determination as the final performance index.

[0032] The model is configured with an early stopping strategy. If the performance metrics on the validation set do not improve for 30 consecutive epochs, the early stopping strategy is triggered, and model training terminates prematurely. The coefficient of determination R0 2 The calculation formula is as follows:

[0033]

[0034] Where y i Represents the phenotypic values ​​of the real samples. This represents the phenotypic value of the predicted sample. The coefficient of determination represents the mean of the true samples. If the final coefficient of determination is equal to 1, the model is considered to have accurately predicted all true values ​​without bias. If the coefficient of determination is less than 1 but greater than 0, it indicates that the model's fit is worse than the mean model. If the coefficient of determination is equal to 0, it indicates that the model's fit is improved compared to the mean model. If the coefficient of determination is less than 0, it indicates that the model's fit is worse than the mean model.

[0035] As a preferred embodiment of the rice phenotypic prediction method based on whole-genome selection described in this invention, the prediction model trained with the best input data includes five convolutional neural network models based on attention mechanisms obtained from a selected set of SNPs, the prediction model trained with the best input data is obtained, the set of SNPs most suitable for the prediction model is found, and the optimal prediction model is obtained; phenotypic prediction is then performed.

[0036] Another objective of this invention is to provide a system for predicting rice phenotypes based on genome-wide selection. This system improves the model's ability to handle complex data and enhances its predictive performance by introducing an attention mechanism and a convolutional neural network. Furthermore, it improves the model's generalization ability and avoids overfitting by employing a five-fold cross-validation and early stopping strategy, thus ensuring the model's accuracy.

[0037] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a rice phenotypic prediction system based on whole-genome selection, comprising a data acquisition module, a genome alignment and variation detection module, a data quality control module, a genetic analysis module, a model training and validation module, and a model selection and prediction module;

[0038] The data acquisition module is used to acquire paired-end sequencing data of rice.

[0039] The genome alignment and variant detection module is used to perform genome alignment and variant detection to obtain SNP data;

[0040] The data quality control module is used for quality control of SNP data;

[0041] The genetic analysis module is used to perform linkage disequilibrium analysis, population structure analysis, and kinship analysis to screen out significant SNP sets;

[0042] The model training and validation module is used to train and validate a convolutional neural network model based on an attention mechanism.

[0043] The model selection and prediction module is used to select the best-performing prediction model and use that model to predict rice phenotypes.

[0044] A computer device includes a memory and a processor, the memory storing a computer program, characterized in that the processor executes the computer program to implement the steps of a rice phenotypic prediction method based on whole-genome selection.

[0045] A computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of a rice phenotypic prediction method based on whole-genome selection.

[0046] The beneficial effects of this invention are as follows: Selecting SNPs that are significantly associated with the phenotype and screened through genome-wide association analysis as input features of the model ensures that the model focuses on features with significant influence while reducing the sample size of features, thus improving model efficiency and training effectiveness. By calculating the attention weight of each genetic variation site, more emphasis and weight are placed on important sites, thereby increasing the model's focus on these important sites and accurately capturing phenotype-related genetic variations. By simultaneously using convolutional kernels of different sizes to extract features at different scales and summing their feature channels, the model can better understand features at different scales and levels of the input data, enhancing the model's ability to abstract genetic data. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0048] Figure 1 A flowchart of a rice phenotypic prediction method based on whole-genome selection provided in one embodiment of the present invention.

[0049] Figure 2 A convolutional neural network framework based on an attention mechanism is shown in an embodiment of the present invention for a rice phenotypic prediction method based on whole-genome selection.

[0050] Figure 3 This is a structural diagram of a rice phenotypic prediction system based on whole-genome selection, provided as an embodiment of the present invention. Detailed Implementation

[0051] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0052] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0053] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0054] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0055] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0056] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0057] Example 1

[0058] Reference Figure 1 and Figure 2As an embodiment of the present invention, a method for predicting rice phenotypes based on genome-wide selection is provided, comprising:

[0059] S1: Obtain paired-end sequencing data of rice, perform genome alignment and variant detection to obtain SNP data, and perform data quality control;

[0060] Furthermore, the genome alignment and variant detection includes comparing sequencing data with a rice reference genome to determine the position of each read on the genome, and performing variant detection on the alignment results to obtain SNP data of rice.

[0061] Quality control was performed by checking the deletion rate of rice samples obtained after alignment with the reference genome. Samples with a missing SNP rate greater than or equal to 0.2 were rejected. Individual information for each SNP in the rice samples was also checked; SNPs with a missing individual information rate greater than or equal to 0.2 were rejected. The minor allele frequency of each SNP in the rice samples was also checked; SNPs with a minor allele frequency less than or equal to 0.05 were rejected. SNPs with high deletion rates and low minor allele frequencies were then removed. Finally, missing SNPs were filled using major alleles.

[0062] The formula for calculating the sample missing rate is as follows:

[0063]

[0064] By calculating the missing rate of each sample, samples with a missing rate greater than or equal to 0.2 are finally screened out.

[0065] The formula for calculating the SNP missing rate is as follows:

[0066]

[0067] By calculating the individual information missing rate for each SNP, SNPs with an individual information missing rate greater than or equal to 0.2 are finally screened out.

[0068] S2: Using the quality-controlled SNP data, linkage disequilibrium analysis, population structure analysis, and kinship analysis were performed to screen out significant SNP sets;

[0069] Furthermore, the linkage disequilibrium analysis includes performing linkage disequilibrium analysis by calculating and evaluating the relationship between linkage haplotype frequencies and SNP sites.

[0070] Among them, the chain disequilibrium value r 2 The calculation is as follows:

[0071] D = P AB -P A PB

[0072]

[0073] Where A is the major allele of SNP1, B is the major allele of SNP2, and P AB P represents the frequency of AB as actually observed. A P B Let r be the expected value of the AB frequency. If a cascading imbalance occurs, the actually observed AB frequency will definitely not be equal to the expected value of the AB frequency. 2 A value of 1 indicates a complete linkage disequilibrium between the two SNPs, while r... 2 A value of 0 indicates a complete linkage equilibrium between the two SNPs. The closer the two SNP sites are, the stronger their linkage relationship. 2 The larger it is.

[0074] The population structure analysis includes using ADMIXTURE software to perform population structure analysis, setting key parameters such as population size, and constructing a genetic correlation matrix. The population size is determined by the number of species contained in the sample.

[0075] The kinship analysis includes using a mixed linear model to perform association analysis on genotype and phenotype, using the genetic correlation matrix and kinship matrix of population structure as covariates to reduce potential confounding effects, and using different common thresholds for genome-wide association analysis (1e-1, 1e-2, 1e-3, 1e-4, 1e-5) to screen the set of SNPs.

[0076] By comparing the results of genome-wide association analysis with linkage disequilibrium, a group of loci showed significant association in the association analysis, and the linkage disequilibrium value r among SNP loci was [value missing]. 2 If all values ​​are greater than or equal to the average attenuation distance, the accuracy of the GWAS analysis results is considered high. The formula for the average attenuation distance is as follows:

[0077]

[0078] The average decay distance is the linkage disequilibrium value r among all SNP sites. 2 Half of the maximum value, where n is the number of pairwise combinations in SNP.

[0079] S3: Divide the selected SNP set into training set, validation set and test set according to the proportion, and use a convolutional neural network model based on attention mechanism for training and validation;

[0080] Furthermore, the training and validation of the attention-based convolutional neural network model includes using the selected SNP set as the input data of the model, dividing it into a training set, a validation set, and a test set in an 8:1:1 ratio, and training the attention-based convolutional neural network prediction model.

[0081] An attention mechanism module is introduced, which weights the SNPs by multiplying the original input with the weights output by the attention module. The attention mechanism allows the model to dynamically learn and selectively focus on different SNPs.

[0082] The importance of each SNP to the prediction task is determined by calculating the attention weight of each SNP. The calculation formula is as follows:

[0083] Output=X*σ(W1(ReLU(W0(X)+b0))+b1)

[0084] The weights of input X are calculated through two fully connected layers, with parameters W0 and b0 and W1 and b1, respectively.

[0085] During the computation, after the first fully connected layer, the ReLU nonlinear activation function is applied to activate the output. This increases the model's nonlinear expressive power. After the second fully connected layer, the Sigmoid function is applied to map the output to the range of 0 to 1, thereby obtaining the weights of X. The product of X and the weights is the final output result.

[0086] It should be noted that this weighting is determined based on the contribution of SNPs to the task; more important SNPs will receive higher weights, while less important SNPs will receive lower weights. This allows the model to focus more on important information, process input data more flexibly and accurately, and improve its predictive performance for rice phenotypes.

[0087] Furthermore, the training and validation of the attention-based convolutional neural network model also includes introducing a residual module to obtain gene features at different levels. The residual module is introduced to obtain features of genotype data at different scales within the model. It allows the model to capture features of genotype data under different receptive fields and fuse these features at different scales into the final feature representation. In this module, the preferred kernel sizes are 20 and 4. Using a convolutional layer with a kernel size of 20 captures larger-scale gene features, while using a convolutional layer with a kernel size of 4 captures smaller-scale gene features. These convolutional layers generate feature maps of different sizes. Then, the gene features at different scales are combined as input to the next layer.

[0088] It should be noted that the fusion method used involves overlaying feature maps at different scales by channel. This allows the model to simultaneously consider gene features at different scales and comprehensively utilize this information to improve the accuracy of rice phenotype prediction. Introducing the residual module enables the model to more comprehensively capture the different scale features of genotype data, thereby enhancing its predictive ability for rice phenotypes.

[0089] S4: Introduce attention mechanism module and residual module, use mean absolute error as loss function, train with stochastic gradient descent optimizer, perform five-fold cross-validation on the model, calculate the mean of the coefficient of determination as the final performance index, and set early stopping strategy to improve the generalization ability of the model.

[0090] Furthermore, the training and validation of the attention-based convolutional neural network model also includes using the mean absolute error as the loss function and training with a stochastic gradient descent optimizer. The specific parameters of the model are as follows: the learning rate is set to 1×10-3.

[0091] It should be noted that a larger learning rate can make the model converge faster, but at the same time it may cause the model to bounce around the optimal point and fail to reach the optimal point. Therefore, in practice, the learning rate can be appropriately reduced or increased according to the model's convergence performance.

[0092] The weight decay coefficient is set to 1×10⁻⁶. It should be noted that this coefficient is used to reduce the size of the weights and prevent overfitting. A larger weight decay coefficient will suppress overfitting but also reduce model performance, while a smaller coefficient will increase the risk of overfitting. Therefore, in practice, it can be appropriately increased or decreased based on the model's fit. The momentum is set to 0.90, which helps the model skip local optima and converge to the global optimum more quickly.

[0093] It should be noted that, typically, momentum can be set between 0.90 and 0.99, and epochs can be set to 200. Larger epochs can lead to overfitting, while smaller epochs can lead to underfitting. Therefore, the actual choice can be adjusted appropriately based on the model's fit. To achieve fast convergence, the training set is divided into multiple batches, with the batch size set to 16. Larger batches can reduce training noise but increase computational resource requirements, while smaller batches can improve the model's generalization ability but increase training noise. The actual choice can be adjusted based on computational resources and the number of samples.

[0094] To more accurately evaluate the model's performance on different subsets of data, avoid over-reliance on results from a single training or validation set, and improve the reliability of the evaluation, a five-fold cross-validation test was performed, and the mean of the five determination coefficients was used as the final performance metric. To track changes in the metric and improve the model's generalization ability, an early stopping strategy was implemented. When the performance metric on the validation set fails to improve for 30 consecutive epochs, the early stopping strategy is triggered, and model training terminates prematurely. The determination coefficient R0 is... 2 The calculation formula is as follows:

[0095]

[0096] Where y i Represents the phenotypic values ​​of the real samples. This represents the phenotypic value of the predicted sample. The coefficient of determination (COP) represents the mean of the true samples. A COP of 1 indicates that the model accurately predicts all true values ​​without bias. A COP less than 1 but greater than 0 indicates that the model fits better than the mean model. A COP of 0 indicates that the model fits close to the mean model. A COP less than 0 indicates that the model fits worse than the mean model.

[0097] S5: Train five attention-based convolutional neural network models using five different SNP datasets, obtain the prediction model trained on the best input data, and perform phenotypic prediction.

[0098] Furthermore, the method of obtaining the prediction model trained with the best input data includes training five different data sets based on attention mechanism using five sets of SNP data to select SNP sets with thresholds (1e-1, 1e-2, 1e-3, 1e-4, 1e-5) for genome-wide association analysis. This results in obtaining the prediction model trained with the best input data. This method can help find the most suitable SNP set for the prediction model based on different data characteristics, thereby obtaining the optimal prediction model.

[0099] The model was input using genome-wide association analysis results, and five thresholds were used to obtain data in five dimensions. Five different data were used as inputs to obtain five prediction models, and the optimal model was selected as the final prediction model.

[0100] It should be noted that predicting phenotypes based on genomic data can further guide breeding. In production, traditional breeding methods typically require years of observation and reproduction because they are based on examining phenotypes and selecting for breeding. However, phenotypic prediction models constructed from genomic data can quickly and accurately predict an individual's phenotype, such as yield, tolerance, and disease resistance, by predicting the association between genotype and phenotype. This allows for earlier screening and selection of individuals with the desired phenotypic characteristics for breeding, thereby accelerating the breeding process.

[0101] Example 2

[0102] The second embodiment of the present invention differs from the first embodiment in that:

[0103] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0104] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0105] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0106] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0107] Example 3

[0108] Reference Figure 3 This is the third embodiment of the present invention, which provides a power plant security intelligent monitoring system, including a data acquisition and quality control module, an analysis module, a data partitioning module, a model training and verification module, a model optimization module, and a prediction module;

[0109] The data acquisition and quality control module is used to acquire paired-end sequencing data of rice, perform genome alignment and variant detection, obtain SNP data, and perform data quality control.

[0110] The analysis module is used to perform linkage disequilibrium analysis, population structure analysis, and kinship analysis using the quality-controlled SNP data to screen out significant SNP sets.

[0111] The data partitioning module is used to divide the selected SNP set into training set, validation set and test set according to the proportion;

[0112] The model training and validation module is used to train and validate a convolutional neural network model based on an attention mechanism.

[0113] The model optimization module is used to introduce the attention mechanism module and the residual module, use the mean absolute error as the loss function, train the model using a stochastic gradient descent optimizer, perform a five-fold cross-validation on the model, calculate the mean of the coefficient of determination as the final performance index, and set an early stopping strategy to improve the model's generalization ability.

[0114] The prediction module is used to train five attention-based convolutional neural network models on five different datasets using five SNP datasets, obtain the prediction model trained on the best input data, and perform phenotypic prediction.

[0115] Example 4

[0116] The fourth embodiment of the present invention provides a method for predicting rice phenotypes based on whole-genome selection.

[0117] This invention was applied to a whole genome sequence of 150 rice phenotypes with seed setting rate, which contained 11,530,898 SNPs, of which 4,828,942 SNPs remained after quality control.

[0118] After threshold screening in genome-wide association analysis, 443,558 SNPs were selected with threshold 1e-1, 34,840 SNPs with threshold 1e-2, 3,906 SNPs with threshold 1e-3, 1,230 SNPs with threshold 1e-4, and 310 SNPs with threshold 1e-5.

[0119] To demonstrate that data selected using thresholds from genome-wide association analysis (GWAS) data are more suitable for phenotypic prediction, prediction models were constructed using traditional genomic selection methods such as Ridge Regression Best Linear Unbiased Prediction, machine learning models such as Random Forest, Extreme Gradient Boosting, Support Vector Regression, and convolutional neural networks with gene attention mechanisms. A window size of 2000 bp and a sliding distance of 2000 bp were used, and in each sliding step, only the linkage disequilibrium association value r between SNPs and other SNPs was retained. 2The goal of selecting SNPs with a threshold of less than 0.01 is to preserve genome-wide characteristics to the greatest extent possible while reducing the number of characteristics. However, the association between genetic variation and phenotype is not considered. Ultimately, 3900 SNPs were selected as input to the model, and the highest coefficient of determination for the optimal model was 0.40. In contrast, using 3906 SNPs selected by genome-wide association analysis with a threshold of 1e-3 as input to the model, the highest coefficient of determination for the optimal model was 0.79. Therefore, using data selected by thresholding in genome-wide association analysis as input to the model can significantly improve the model's performance.

[0120] To demonstrate that attention-based convolutional neural networks can better predict phenotypes, we used five different thresholds to filter data and constructed prediction models using the traditional genomic selection method Ridge Regression Best Linear Unbiased Prediction, the machine learning models Random Forest, Extreme Gradient Boosting, Support Vector Regression, and attention-based convolutional neural networks. The results showed that in Ridge Regression Best Linear Unbiased Prediction, the optimal dataset was the SNP set filtered to the power of 1e-3, with a model determination coefficient of 0.73; in Random Forest, the optimal dataset was 1e-3, with a model determination coefficient of 0.69; in Extreme Gradient Boosting, the optimal dataset was 1e-2, with a model determination coefficient of 0.57; in Support Vector Regression, the optimal dataset was 1e-3, with a model determination coefficient of 0.77; and in attention-based convolutional neural network models, the optimal dataset was 1e-3, with a model determination coefficient of 0.79. Therefore, convolutional neural network models based on attention mechanisms are better able to capture the correlation between variant features and phenotypes.

[0121] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting phenotypes of rice based on whole genome selection, characterized by, include: Rice paired-end sequencing data were obtained, genome alignment and variant detection were performed to obtain SNP data, and data quality control was carried out. Using the quality-controlled SNP data, linkage disequilibrium analysis, population structure analysis, and kinship analysis were performed to screen out significant SNP sets; The selected SNP set was divided into training set, validation set and test set according to the proportion, and a convolutional neural network model based on attention mechanism was used for training and validation. An attention mechanism module and a residual module are introduced. The mean absolute error is used as the loss function, and the model is trained using a stochastic gradient descent optimizer. The model is subjected to a five-fold cross-validation test, and the mean of the coefficient of determination is calculated as the final performance index. An early stopping strategy is set to improve the generalization ability of the model. Using five SNP datasets, five attention-based convolutional neural network models were trained on different datasets to obtain the prediction model trained on the best input data for phenotypic prediction. The linkage disequilibrium analysis includes performing linkage disequilibrium analysis by calculating and evaluating the relationship between linkage haplotype frequencies and SNP sites; Among them, the value of chain disequilibrium relationship The calculation is expressed as: = Where A is the major allele of SNP1, and B is the major allele of SNP2. The frequency of AB as actually observed. This represents the expected value of the AB frequency. when A value of 1 indicates that the two SNPs have a complete linkage disequilibrium relationship. When the value is 0, it indicates that there is a complete linkage equilibrium between the two SNPs; The population structure analysis includes using ADMIXTURE software to perform population structure analysis, setting key parameters such as population size and constructing a genetic correlation matrix. The population size is determined by the number of species contained in the sample. The kinship analysis includes using a mixed linear model to perform association analysis on genotype and phenotype, using the genetic correlation matrix and kinship matrix of population structure as covariates to reduce potential confounding effects, and using different genome-wide association analysis thresholds to screen the set of SNPs; By comparing the results of genome-wide association analysis with linkage disequilibrium, if a group of loci show significant association in the association analysis, and the linkage disequilibrium values ​​between SNP loci are high, then the association is considered successful. All values ​​are greater than or equal to the average attenuation distance, indicating a high degree of accuracy in the GWAS analysis results; The average attenuation distance is expressed as: The average decay distance represents the linkage disequilibrium value among all SNP sites. Half of the maximum value, where n is the number of pairwise combinations of SNPs; The training and validation of the attention-based convolutional neural network model includes using the selected SNP set as the input data of the model, dividing it into a training set, a validation set, and a test set in an 8:1:1 ratio, and training the attention-based convolutional neural network prediction model. An attention mechanism module is introduced, which weights the SNPs by multiplying the original input with the weights output by the attention module. The attention mechanism allows the model to dynamically learn and selectively focus on different SNPs. The importance of each SNP to the prediction task is determined by calculating the attention weight of each SNP. The calculation formula is as follows: Out Where input The weights are calculated through two fully connected layers, the parameters of which are respectively... and as well as and ; The training and validation of the attention-based convolutional neural network model also includes introducing a residual module to obtain features of different scales of genotype data, using a convolutional layer with a kernel size of 20 to capture larger-scale gene features, using a convolutional layer with a kernel size of 4 to capture smaller-scale gene features, generating feature maps of different sizes, combining gene features of different scales as input to the next layer, and stacking and fusing feature maps of different scales by channel.

2. The rice phenotypic prediction method based on genome-wide selection as described in claim 1, characterized in that: The genome alignment and variant detection process includes comparing sequencing data with a rice reference genome to determine the position of each read in the genome, performing variant detection on the alignment results to obtain SNP data for rice, and performing quality control by checking the missing rate of rice samples after alignment with the reference genome. If the missing rate of a sample is greater than or equal to 0.2%, the sample is removed. For each SNP in the rice sample, we checked whether it lacked individual information. If the rate of missing individual information for a single SNP was greater than or equal to 0.2, it was directly screened out. We also checked the minor allele frequency for each SNP in the rice sample. If the minor allele frequency for a single SNP was less than or equal to 0.05, it was directly screened out. We screened out SNPs with high deletion rates and low minor allele frequencies, and finally filled the missing SNPs with major alleles.

3. The rice phenotypic prediction method based on genome-wide selection as described in claim 2, characterized in that: The training and validation of the attention-based convolutional neural network model also includes using the mean absolute error as the loss function, training with a stochastic gradient descent optimizer, and setting the learning rate to [value missing]. The weight decay coefficient is set to The momentum was set to 0.90, the epoch was set to 200, the model was subjected to a five-fold cross-validation test, and the mean of the five coefficients of determination was taken as the final performance index. The model is configured with an early stopping strategy. If the performance metrics on the validation set do not improve for 30 consecutive epochs, the early stopping strategy is triggered, and the model training terminates prematurely, reducing the coefficient of determination. The calculation formula is as follows: in Represents the phenotypic values ​​of the real samples. This represents the phenotypic value of the predicted sample. The coefficient of determination represents the mean of the true samples. If the final coefficient of determination is equal to 1, the model is considered to have accurately predicted all true values ​​without bias. If the coefficient of determination is less than 1 but greater than 0, it indicates that the model fits better than the mean model. If the coefficient of determination is equal to 0, it indicates that the model fits better than the mean model. If the coefficient of determination is less than 0, it indicates that the model fits the model better than the mean model.

4. The rice phenotypic prediction method based on genome-wide selection as described in claim 3, characterized in that: The prediction model for obtaining the best input data training includes five attention-based convolutional neural network models obtained from the selected SNP set, obtaining the prediction model for the best input data training, finding the SNP set most suitable for the prediction model, and obtaining the optimal prediction model.

5. A system employing the rice phenotypic prediction method based on whole-genome selection as described in any one of claims 1 to 4, characterized in that: It includes modules for data acquisition and quality control, analysis, data partitioning, model training and validation, model optimization, and prediction. The data acquisition and quality control module is used to acquire paired-end sequencing data of rice, perform genome alignment and variant detection, obtain SNP data, and perform data quality control. The analysis module is used to perform linkage disequilibrium analysis, population structure analysis, and kinship analysis using the quality-controlled SNP data to screen out significant SNP sets. The data partitioning module is used to divide the selected SNP set into training set, validation set and test set according to the proportion; The model training and validation module is used to train and validate a convolutional neural network model based on an attention mechanism. The model optimization module is used to introduce the attention mechanism module and the residual module, use the mean absolute error as the loss function, train the model using a stochastic gradient descent optimizer, perform a five-fold cross-validation on the model, calculate the mean of the coefficient of determination as the final performance index, and set an early stopping strategy to improve the model's generalization ability. The prediction module is used to train five attention-based convolutional neural network models on five different datasets using five SNP datasets, obtain the prediction model trained on the best input data, and perform phenotypic prediction.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

Images