Plant phenotype prediction

Abstract

The present invention provides a method and apparatus for predicting plant phenotypes. According to one example of the method, graph-structured data for each plant is constructed using multi-omics data, such as genome, transcriptome, and metabolome, as graph nodes, and the correlations between different omics as graph edges. The constructed graph-structured data is then input into a graph convolutional neural network to extract node features, which are then updated using a Transformer network. The node features are then combined and input into a fully connected layer to output a predicted phenotype. In this way, the graph convolutional neural network is combined with the Transformer network to achieve gene-to-phenotype prediction, and multi-omics data is used to construct a graph structure and integrate the multi-omics data to achieve highly accurate phenotype prediction, thereby effectively improving the effectiveness of phenotype prediction.

Description

[Technical Field]

[0001] The present invention relates to the fields of deep learning technology and bioinformatics, and in particular to plant phenotype prediction. [Background technology]

[0002] Gene-to-phenotype prediction plays an important role in identifying causative genes for disease prediction and predicting crop yield, making it a central challenge in genetic biology. Currently, phenotype prediction is based on genome analysis. However, phenotype expression structure is related not only to genome analysis but also to multi-omics analysis, such as transcriptome and metabolome analysis. Direct gene-to-phenotype prediction has a too large span, resulting in poor phenotype prediction results. Therefore, to achieve accurate phenotype prediction, it is necessary to perform phenotype prediction by linking multi-omics. Furthermore, a graph structure can associate multiple types of nodes, and graph neural networks can be used to mine the association relationships between multiple nodes, enabling multi-omics integration for phenotype prediction.

[0003] In recent years, Transformers have been widely used in computer vision, and their attention mechanism can learn the associations between different local features. In particular, there have been related studies applying Transformers to protein coding region prediction. However, there have been few studies applying Transformers to phenotype prediction. Summary of the Invention

[0004] Therefore, the present invention provides a phenotype prediction method using a combination of a graph convolutional neural network and a Transformer network, which can integrate multi-omics data to achieve phenotype prediction and effectively solve the problem of poor phenotype prediction using a single omics. In particular, the present invention provides a method and apparatus for plant phenotype prediction, in which multi-omics graph structure data is constructed and correlations between multi-omics are extracted using a method combining a graph convolutional neural network and a Transformer network, thereby achieving multi-omics phenotype prediction.

[0005] According to a first aspect of an embodiment of the present invention, (1) obtaining multi-omics data and phenotype data of a target species, performing quality control and dimension reduction processing on the multi-omics data to obtain multiple omics features, and performing outlier and missing value processing on the phenotype data to obtain true phenotype data; (2) performing a normalization process on the multiple omics features obtained in step (1), calculating a first similarity between the multiple omics features of each sample of the target species, and constructing graph structure data of each sample using the first similarity; (3) for each sample, inputting the graph structure data of each sample obtained in step (2) into a three-layer graph neural network to extract features of each node of each sample; (4) for each sample, the features of each node of each sample extracted in step (3) are combined in the order of multi-omics, and then input into two first fully connected layers to obtain a first phenotype prediction value, and calculate a first loss based on the first phenotype prediction value and the corresponding true phenotype data; (5) for each sample, inputting the features of each node of each sample obtained in step (3) as local identifier embedding features into a Transformer network together with the global identifier embedding feature to obtain updated global identifier embedding features and local identifier embedding features of each node, calculating a second similarity between the updated global identifier embedding feature and the local identifier embedding feature of each node, combining the local identifier embedding features and the global identifier embedding features of the two nodes with the highest second similarity, and then inputting them into two second fully connected layers to obtain a second phenotype prediction value, and calculating a second loss based on the second phenotype prediction value and corresponding true phenotype data; (6) repeating steps (3) to (5) to perform training, and adjusting network parameters of the graph neural network, the first fully connected layer, the Transformer network, and the second fully connected layer using backpropagation algorithm based on a total loss obtained by adding the first loss and the second loss, to obtain the trained graph neural network, the first fully connected layer, the Transformer network, and the second fully connected layer; (7) A method for predicting a plant phenotype is provided, comprising the steps of: obtaining a first phenotype prediction value for a material to be detected using the trained graph neural network and the first fully connected layer; obtaining a second phenotype prediction value using the trained graph neural network, the Transformer network, and the second fully connected layer; and determining an average value of the first phenotype prediction value and the second phenotype prediction value as a final predicted phenotype result for the material to be detected.

[0006] Furthermore, the multi-omics data includes genome data, transcriptome data, and metabolome data, and the graph neural network includes a graph convolution layer and a ReLU activation layer.

[0007] Furthermore, in step (1), the step of performing quality control and dimension reduction processing on the multi-omics data to obtain a plurality of omics features and performing outlier and missing value processing on the phenotype data to obtain true phenotype data includes: (1.1) removing missing values ​​and abnormal outliers in the phenotypic data of each sample of the target species; (1.2) completing quality control filtering for the multi-omics data by removing locations with a missing rate of more than 90% in the multi-omics data of each sample of the target species and removing loci with a minor allele frequency of less than 5% in the genomic data; (1.3) encoding the quality control filtered multi-omics data in step (1.2), wherein the genomic data is encoded as -1, 0, or 1 according to homozygous, heterozygous, or homozygous variant, respectively, and the other omics data is not encoded, and its measured value is used; (1.4) For each position of each omics data, calculate the Pearson correlation coefficient between the encoded features of all samples and the corresponding phenotype data to obtain the degree of association between each position of each omics data and the phenotype data, and select L positions with the highest degree of association for each omics data as omics features to obtain multiple omics features.

[0008] Furthermore, the formula for calculating the Pearson correlation coefficient is:

number

[0009] Furthermore, step (2) is (2.1) A step of performing normalization processing on each omics feature obtained in step (1.4), specifically, there are a total of M samples, each sample has N types of omics, and the j-th dimension-reduced omics feature of the i-th sample is defined as fij Let i=1...M, j=1...N.

number

number

[0010] Furthermore, the first loss includes a predicted regression loss and a Pearson correlation coefficient loss, the formula of which is:

number

number

number

[0011] Furthermore, step (5) (5.1) A step of generating a global identifier embedding feature by initializing using random normal distribution parameters, where the values ​​of each dimension of the global identifier embedding feature are all taken from random values ​​generated by normal distribution, and the feature length is the same as the feature length of each node extracted in step (3); (5.2) for each sample, calculate a first attention matrix for the sample based on the features of all nodes extracted in step (3); (5.3) for each sample, the features of each node extracted in step (3) are used as local identifier embedding features, and the global identifier embedding features obtained in step (5.1) and the local identifier embedding features of each node are input into a Transformer network including a self-attention module and a feed-forward neural network module; the self-attention module calculates a second attention matrix corresponding to the sample; the second attention matrix and the first attention matrix obtained in step (5.2) are combined to update the global identifier embedding features and the local identifier embedding features of each node; and the feed-forward neural network module obtains the updated global identifier embedding features and the local identifier embedding features of each node. (5.4) Calculating the cosine similarity between the updated global identifier embedding feature obtained in step (5.3) and the local identifier embedding feature of each node as a second similarity, combining the local identifier embedding feature and the global identifier embedding feature of the two nodes with the highest second similarity to obtain a second combined feature, inputting the second combined feature into two second fully connected layers to obtain a second phenotype prediction value, and calculating a second loss based on the second phenotype prediction value and corresponding true phenotype data.

[0012] Furthermore, the second loss is the predicted regression loss, whose formula is:

number

[0013] According to a second aspect of the present invention, there is provided a plant phenotype prediction device including one or more processors and a memory, the memory being connected to the processor and used to store program data, and the processor being used to execute the program data to implement the plant phenotype prediction method.

[0014] According to a third aspect of the present invention, there is provided a computer-readable storage medium storing a program, which, when executed by a processor, performs the plant phenotype prediction method.

[0015] By using multi-omics as nodes, constructing a graph structure based on the correlations between multi-omics, and using a graph neural network to perform regression tasks across the entire graph, the graph neural network is combined with a Transformer to mine correlations between different omics, integrating the attention of the graph nodes into the attention update of the Transformer, and dynamically selecting omics features that contribute more to phenotype prediction based on their degree of correlation with global features, thereby effectively solving the problem of low accuracy in single-omics phenotype prediction. In other words, the present invention creatively uses a method combining a graph neural network and a Transformer to achieve multi-omics phenotype prediction, which is advantageous for improving the effectiveness of phenotype prediction. [Brief explanation of the drawings]

[0016] [Figure 1] FIG. 1 is a detailed diagram of the structure of the multi-omics-based phenotype prediction network of the present invention. [Figure 2] 1 is a flowchart of the plant phenotype prediction method of the present invention. [Figure 3] 1 is a test flowchart of the plant phenotype prediction method of the present invention. [Figure 4] FIG. 10 is a comparison diagram of experimental results of the present invention. [Figure 5] 1 is a schematic diagram showing the structure of a plant phenotype prediction device of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0017] Illustrative embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description refers to the drawings, identical numerals in different drawings refer to the same or similar elements unless otherwise noted. The embodiments described in the following illustrative examples do not represent all embodiments consistent with the present invention. On the contrary, they are merely examples of apparatus and methods consistent with certain aspects of the present invention, as set forth in the appended claims.

[0018] The terms used in the present invention are merely for the purpose of describing particular embodiments and are not intended to limit the present invention. As used in the present invention and the appended claims, the singular forms "a," "the," and "the" are intended to include the plural forms unless the context clearly indicates otherwise. Also, the term "and / or" as used in the present invention should be understood to mean and include any and all possible combinations of one or more of the associated listed items.

[0019] While the present invention may use terms such as first, second, and third to describe various pieces of information, it should be understood that such information is not limited to these terms. These terms are used only to distinguish between the same types of information. For example, first information may be referred to as second information, and similarly, second information may be referred to as first information, without departing from the scope of the present invention. Depending on the context, the term "if..." as used herein may be interpreted as "with...," "when...," or "in response to a determination."

[0020] The present invention will now be described in detail with reference to the accompanying drawings. Unless inconsistent, the features in the following examples and embodiments can be combined with each other.

[0021] Referring to FIG. 2, the plant phenotype prediction method of the present invention specifically includes the following steps:

[0022] (1) Multi-omics data and phenotype data of the target species are obtained, quality control and dimensionality reduction are performed on the multi-omics data to obtain multiple omics features, and outlier and missing value processing is performed on the phenotype data to obtain the true phenotype data.

[0023] Furthermore, multi-omics data includes, but is not limited to, genome data, transcriptome data, and metabolome data, i.e., multi-omics includes, but is not limited to, genome, transcriptome, and metabolome, etc.

[0024] In this example, an open-source tomato dataset was used to obtain multi-omics and phenotypic data for the target species. The open-source tomato dataset included a total of 332 tomato samples, each of which contained 6,971,059 SNPs (single nucleotide polymorphisms), 657,549 InDels (insertion and deletion), 54,838 SVs (structural variations), and approximately 170,000 RAN-Seg gene expression data.

[0025] Hereinafter, a method for performing quality control and dimension reduction processing on multi-omics data and outlier and missing value processing on phenotype data will be described in detail, taking the example of obtaining multi-omics data and phenotype data of a target species from an open-source tomato dataset. In particular, step (1) may specifically include the following substeps: (1.1) Remove missing values ​​and abnormal outliers in the phenotypic data of each sample of the target species in the tomato dataset. (1.2) Quality control filtering of the multi-omics data is completed by removing positions with a missing rate of more than 90% for each position in the multi-omics data of each sample of the target species in the tomato dataset and removing loci with a minor allele frequency of less than 5% in the genomic data. (1.3) Encode the quality-controlled filtered multi-omics data in step (1.2), and the genomic data are analyzed for homozygotes (homozygotes). , e.g. aa ), heterozygote, homozygous mutant (other homozygote , e.g., AA ) are coded as -1, 0, or 1, respectively, and other omics data are not coded and their measured values ​​are used. (1.4) For each position of each omics data, calculate the Pearson correlation coefficient between the coding features of all samples and the corresponding phenotype data to obtain the degree of association between each position of each omics data and the phenotype data, sort them in descending order of association, and select the previous L positions for each omics data as the omics feature corresponding to that omics data to obtain multiple omics features.

[0026] Note that for each omics data, selecting the L most highly correlated positions as the corresponding omics features is equivalent to performing dimensionality reduction on the omics data. In this example, the tomato data contains four omics, resulting in four omics features for each plant sample, which can be collectively represented by a 4*L omics feature. The value of L can be set according to the actual needs of the application scenario. For example, in this example, the first L = 1000 positions for each omics data are selected as the omics features after dimensionality reduction.

[0027] Furthermore, the formula for calculating the Pearson correlation coefficient is:

number

[0028] where ρ XY is the Pearson correlation coefficient, X and Y represent the coded features and the corresponding phenotype data, respectively, cov(*) represents the covariance, and σ represents the standard deviation. XY is the Pearson correlation coefficient calculated by the above formula, and the value of this coefficient ranges from -1 to 1. A Pearson correlation coefficient smaller than 0 indicates a negative correlation, a Pearson correlation coefficient greater than 0 indicates a positive correlation, and a Pearson correlation coefficient equal to 0 indicates no correlation.

[0029] The larger the absolute value of the Pearson correlation coefficient, the closer the correlation between the two variables, i.e., the higher the degree of association. Thus, based on the Pearson correlation coefficient, the degree of association between each position of each omics data and the phenotype data can be obtained.

[0030] (2) Normalizing the omics features obtained in step (1), calculating the first similarity between the multiple omics features of each sample of the target species, and constructing graph structure data for each sample using the first similarity, where nodes in the graph represent different omics, and edges in the graph represent the first similarity between different omics.

[0031] (2.1) Step (1) The normalization process is performed on each omics feature obtained in step 1. The specific method is as follows: There are a total of M samples, each of which has N types of omics. The j-th dimension-reduced omics feature of the i-th sample is defined as f ij Let i=1...M, j=1...N.

number

[0032] where f ij ′ represents the j-th omics feature after normalization of the i-th sample, and u j is the average value of all samples of the jth omics, and var j is the standard deviation of all samples of the jth omics.

[0033] Note that each sample has N types of omics, that is, each sample, that is, each plant, has N types of omics data. Illustratively, in this example, M=332 and N=4.

[0034] (2.2) For each sample, calculate the cosine similarity between the different omics features after the normalization process calculated in step (2.1) as the first similarity, and the calculation formula is:

number

[0035] where similarity is the cosine similarity, A and B represent the omics features after normalization, A·B is the dot product, and ||A|| represents the norm corresponding to the omics feature A after normalization.

[0036] The similarity value ranges from -1 to 1, and a larger similarity value indicates a higher feature similarity and a higher degree of association.

[0037] (2.3) For each sample, construct graph structure data for each sample by taking the different omics of the sample as graph nodes and the first similarity between the different omics features obtained in step (2.2) as the weights of the edges of the graph.

[0038] The weight of the edge reflects the degree of association between different omics, that is, the larger the first similarity, the higher the degree of association between different omics.

[0039] (3) For each sample, the graph structure data obtained in step (2) is input into a three-layer graph neural network to extract the features of each node, and the graph neural network includes a graph convolution layer and a ReLU activation layer.

[0040] Specifically, in the three-layer graph neural network, there is one ReLU activation layer after each graph convolution layer, and the input and output feature dimensions of the three graph convolution layers are all set to, for example, 1000. As shown in Figure 1, the graph structure data of each sample can be input into the three-layer graph neural network to extract the features of each node.

[0041] (4) For each sample, the features of each node of the sample extracted in step (3) are combined in multi-omics order to obtain a first combined feature, and the first combined feature is input into two first fully connected layers to obtain a first phenotype prediction value, and a first loss is calculated based on the first phenotype prediction value and the corresponding true phenotype data.

[0042] The true phenotype data is the phenotype data after processing of outliers and missing values ​​in step (1).

[0043] Specifically, in this embodiment, using the open-source tomato dataset as an example, for each sample, the features of each node extracted in step (3) are combined in order with four types of omics: SNPs, InDels, SVs, and RAN-Seg, to obtain a 4000-dimensional first combined feature. Next, as shown in Figure 1, the 4000-dimensional first combined feature is input to two first fully connected layers to output a first phenotype prediction value. Here, the input dimension of the first first fully connected layer is 4000 and the output dimension is 32, and the input dimension of the second first fully connected layer is 32 and the output dimension is 1. Finally, a first loss is calculated based on the first phenotype prediction value and the corresponding true phenotype data.

[0044] In particular, the first loss may include two parts, which are the predicted regression loss and the Pearson correlation coefficient loss, and its formula is specifically as follows:

number

[0045] where Loss1 represents the first loss, M is the number of samples, which is 332 in this example, and Y i is the true phenotype data of node sample i, and Y i ′ is the first phenotype prediction value of node sample i,

number

number

[0046] (5) For each sample, the features of each node of the sample extracted in step (3) are input as local identifier embedding features into a Transformer network together with the global identifier embedding feature to obtain the updated global identifier embedding feature and the local identifier embedding feature of each node, calculate the second similarity between the global identifier embedding feature and the local identifier embedding feature of each node, combine the local identifier embedding features and global identifier embedding features of the two nodes with the highest second similarity, and then input them into two second fully connected layers to obtain a second phenotype prediction value, and calculate a second loss based on the second phenotype prediction value and the corresponding true phenotype data.

[0047] (5.1) A global identifier embedding feature may be generated by initializing using random normal distribution parameters. Specifically, the values ​​of each dimension of the global identifier embedding feature are taken from random values ​​generated by normal distribution, and the feature length of the global identifier embedding feature is the same as the feature length of each node extracted in step (3).

[0048] Specifically, we obtain a random normal distribution using the xavier_normal function in the Pytorch framework and initialize it using the random normal distribution parameters to generate a global identifier embedding feature. Here, the feature length of the global identifier embedding feature is the same as the feature length of each node extracted in step (3), i.e., the value of the feature length of the global identifier embedding feature is 1000.

[0049] (5.2) For each sample, calculate the first attention matrix of the sample based on the features of all nodes of the sample extracted in step (3), and the calculation formula is:

number

[0050] Here, Attention(F i,F i ) represents the first attention matrix of the i-th sample, softmax(·) represents the softmax activation function, and F i represents the feature matrix of the sample, which consists of the features of all the nodes of the sample obtained in step (3), and the feature vector of each node is F ij In this embodiment, j=1, 2, 3, 4.

[0051] (5.3) For each sample, the feature F of each node of the sample extracted in step (3) ij Let G be the local identifier embedding feature, and the global identifier embedding feature G obtained in step (5.1) and the local identifier embedding feature of each node are input into a Transformer network. Here, the Transformer network includes a self-attention module and a feed-forward neural network module. In this way, the self-attention module calculates a second attention matrix corresponding to the sample, and the second attention matrix and the first attention matrix obtained in step (5.2) are combined to update the feature, and the feed-forward neural network module obtains the updated global identifier embedding feature and the local identifier embedding feature of each node.

[0052] The formula for calculating the second attention matrix is ​​as follows:

number

[0053] where Attention(Q,K) represents the second attention matrix of the sample, Q and K are feature matrices consisting of the global identifier embedding feature and the local identifier feature of each node, respectively, and d k is the feature dimension of the feature matrix Q,K.

[0054] The updated global identifier embedding feature and the local identifier embedding feature of each node can be expressed as follows:

number

[0055] Here, V′ represents the updated global identifier embedding feature and the local identifier embedding feature of each node, and V represents the updated global identifier embedding feature G and the local identifier embedding feature F of each node. ij Represents.

[0056] (5.4) For each sample, calculate the cosine similarity between the updated global identifier embedding feature obtained in step (5.3) and the local identifier embedding feature of each node as the second similarity, combine the local identifier embedding feature and global identifier embedding feature of the two nodes with the highest second similarity to obtain a second combined feature, input the second combined feature into two second fully connected layers to obtain a second phenotype prediction value, and calculate a second loss based on the second phenotype prediction value and the corresponding true phenotype data.

[0057] Specifically, calculating the second similarity between each omics feature and the global feature involves calculating the cosine similarity between the updated global identifier embedding feature obtained in step (5.3) and the local identifier embedding feature of each node, where the cosine similarity is the second similarity. The local identifier embedding feature and the global identifier embedding feature of the two nodes with the highest second similarity are combined to obtain a 3000-dimensional second combined feature. Here, the combining method is the same as the combining method in step (4). Next, as shown in FIG. 1, the second combined feature is input to two second fully connected layers to obtain a second phenotype prediction value. The first second fully connected layer has an input dimension of 3000 and an output dimension of 32, while the second second fully connected layer has an input dimension of 32 and an output dimension of 1. Finally, a second loss is calculated based on the second phenotype prediction value and the corresponding true phenotype data. In particular, the second loss may be a predictive regression loss, and its formula is:

number

[0058] where Loss2 is the second loss, M is the number of samples, which is 332 in this example, and Y i is the true phenotype data of node sample i, and Y i '' is the second phenotype prediction value of node sample i, obtained by prediction of the Transformer network.

[0059] (6) Steps (3) to (5) are repeated to perform training, and based on the total loss obtained by adding the first loss and the second loss, backpropagation using the stochastic gradient descent method is used to adjust the network parameters of the three-layer graph neural network in step (3), the two first fully connected layers in step (4), and the Transformer network and two second fully connected layers in step (5), to obtain a trained three-layer graph neural network and two first fully connected layers, and a Transformer network and two second fully connected layers.

[0060] The training is performed by repeating steps (3) to (5), and the training is stopped when the set number of repetitions is reached, and the final network parameters are obtained. For example, in this embodiment, the number of repetitions may be set to 100.

[0061] (7) For the material to be detected, a first phenotype prediction value is obtained by the trained three-layer graph neural network and two first fully connected layers, and a second phenotype prediction value is obtained by the trained three-layer graph neural network, a Transformer network, and two second fully connected layers, and the average value of the first phenotype prediction value and the second phenotype prediction value is used as the final predicted phenotype result of the material to be detected.

[0062] In this embodiment, the process of predicting a phenotype using a trained three-layer graph neural network with two first fully connected layers and a Transformer network with two second fully connected layers may specifically include the following steps, as shown in FIG. 3 : (7.1) Collect multi-omics data of the material to be detected: SNPs (single nucleotide polymorphisms), InDels (insertion / deletion of bases), SVs (structural variations), and RAN-Seg data. (7.2) Complete quality control filtering for the multi-omics data by removing positions with a missing rate greater than 90% for each position in the multi-omics data of each sample and removing loci with a minor allele frequency less than 5% in the genomic data. (7.3) Quality control: The filtered multi-omics data is coded. Genomic data is coded as -1, 0, or 1 according to the homozygous, heterozygous, or homozygous variant type, respectively. Other omics data is not coded, and its measured value is used. Then, the quality control process is repeated in step (1). 4 ) and perform dimension reduction on each omics data according to the L positions selected for each omics data in (1), and construct graph structure data according to the method in step (2). (7.4) The graph structure data is input to a trained three-layer graph neural network with two first fully connected layers, and a Transformer network with two second fully connected layers to predict a phenotype, the network parameters of which are the network parameters obtained by training in step (6), and the predicted first phenotype prediction value and second phenotype prediction value are obtained, respectively, and the average value of the first phenotype prediction value and the second phenotype prediction value is used as the final predicted phenotype result of the material to be detected.

[0063] The experimental results for the tomato dataset in this example are shown in Table 1, of which SNPs (single nucleotide polymorphisms), InDels (insertion and deletion), SVs (structural variations), and RAN-Seg gene expression are the results of the conventional method RRBLUP. In this example, the results of each Pearson correlation coefficient are shown in Figure 4, and the method described in this invention improved by 11.94% to 84.02% compared to the conventional method. Table 1 Experimental results [Table 1]

[0064] Referring to Figure 5, a plant phenotype prediction device provided by an embodiment of the present invention includes one or more processors and a memory, the memory is connected to the processor, the memory is used to store program data, and the processor is used to execute the program data to implement the plant phenotype prediction method in the above embodiment.

[0065] An embodiment of the plant phenotype prediction device of the present invention can be applied to any device having data processing capabilities, which may be a computer-like device or apparatus. An embodiment of the device may be implemented using software, hardware, or a combination of hardware and software. For example, in a software implementation, the device is formed by a processor in the device having data processing capabilities loading corresponding computer program instructions from a non-volatile memory into an internal memory and executing them. At the hardware level, FIG. 5 is a hardware structural diagram of a device having data processing capabilities in which the plant phenotype prediction device of the present invention is located. In addition to the processor, internal memory, network interface, and non-volatile memory shown in FIG. 5, the device having data processing capabilities in which the embodiment of the device is located may also include other hardware depending on the actual function of the device having data processing capabilities, but these will not be described here.

[0066] The process of realizing the functions and roles of each unit in the above device is specifically detailed in the process of realizing the corresponding step in the above method, so the description will be omitted here.

[0067] The device embodiments basically correspond to the method embodiments, so please refer to the description of some of the method embodiments for relevant parts. The above-mentioned device embodiments are merely exemplary, and units described as separate components may or may not be physically separated, and components shown as units may or may not be physical units, i.e., they may be located in one place or distributed across multiple network units. To achieve the objectives of the present invention, some or all of these modules may be selected according to actual needs. Those skilled in the art can understand and implement the present invention without any creative effort.

[0068] An embodiment of the present invention further provides a computer-readable storage medium storing a program, which, when executed by a processor, performs the plant phenotype prediction method of the above embodiment.

[0069] The computer-readable storage medium may be an internal storage unit such as a hard disk or internal memory of any device having data processing capabilities described in any of the above embodiments. The computer-readable storage medium may also be any device having data processing capabilities, such as a plug-in hard disk, a Smart Media Card (SMC), an SD card, or a flash card, provided in the device. Furthermore, the computer-readable storage medium may include both an internal storage unit and an external storage device of any device having data processing capabilities. The computer-readable storage medium may be used to store computer programs and other programs and data required by any device having data processing capabilities, and may also be used to temporarily store output data or data to be output.

[0070] The above examples are only for explaining the technical solutions of the present invention, and are not intended to limit the same. The present invention has been described in detail with reference to the above examples. However, those skilled in the art should understand that the technical solutions described in the above examples can still be modified or equivalent substitutions can be made for some of the technical features therein, and such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. (1) obtaining multi-omics data and phenotype data of a target species, performing quality control and dimension reduction processing on the multi-omics data to obtain multiple omics features, and performing outlier and missing value processing on the phenotype data to obtain true phenotype data; (2) performing a normalization process on the omics features obtained in step (1), calculating a first similarity between the multiple omics features of each sample of the target species, and constructing graph structure data of each sample using the first similarity; (3) for each sample, inputting the graph structure data obtained in step (2) into a three-layer graph neural network to extract features of each node; (4) for each sample, the features of each node of the sample extracted in step (3) are combined in the order of multi-omics, and then input into two first fully connected layers to obtain a first phenotype prediction value, and calculate a first loss based on the first phenotype prediction value and the corresponding true phenotype data; (5) for each sample, inputting the features of each node of the sample obtained in step (3) as local identifier embedding features together with the global identifier embedding feature into the Transformer network to obtain updated global identifier embedding features and local identifier embedding features of each node, calculating a second similarity between the updated global identifier embedding feature and the local identifier embedding feature of each node, combining the local identifier embedding features and global identifier embedding features of the two nodes with the highest second similarity, and then inputting them into two second fully connected layers to obtain a second phenotype prediction value, and calculating a second loss based on the second phenotype prediction value and the corresponding true phenotype data; (6) repeating steps (3) to (5) to perform training, and adjusting network parameters of the graph neural network, the first fully connected layer, the Transformer network, and the second fully connected layer using a backpropagation algorithm based on a total loss obtained by adding the first loss and the second loss, to obtain the trained graph neural network, the first fully connected layer, the Transformer network, and the second fully connected layer; (7) obtaining a first phenotype prediction value for a material to be detected using the trained graph neural network and the first fully connected layer, obtaining a second phenotype prediction value using the trained graph neural network, the Transformer network, and the second fully connected layer, and setting an average value of the first phenotype prediction value and the second phenotype prediction value as a final predicted phenotype result for the material to be detected. A method for predicting a plant phenotype, comprising:

2. the multi-omics data includes genome data, transcriptome data, and metabolome data; The graph neural network includes a graph convolution layer and a ReLU activation layer. The plant phenotype prediction method according to claim 1 .

3. In step (1), the step of performing quality control and dimension reduction processing on the multi-omics data to obtain a plurality of omics features and performing outlier and missing value processing on the phenotype data to obtain true phenotype data includes: (1.1) Removing missing values ​​and abnormal outliers in the phenotypic data of each sample of the target species; (1.2) completing quality control filtering for the multi-omics data by removing locations with a missing rate of more than 90% in the multi-omics data of each sample of the target species and removing loci with a minor allele frequency of less than 5% in the genomic data; (1.3) encoding the quality-controlled filtered multi-omics data in step (1.2), wherein the genomic data is encoded as -1, 0, 1 according to the homozygous type, heterozygous type, and homozygous variant type, respectively, and the other omics data is not encoded, and its measured value is used; (1.4) For each position of each omics data, calculate the Pearson correlation coefficient between the encoded features of all samples and the corresponding phenotype data to obtain a degree of association between each position of each omics data and the phenotype data, and select L positions with the highest degree of association for each omics data as omics features to obtain a plurality of omics features. The plant phenotype prediction method according to claim 1 .

4. The formula for calculating the Pearson correlation coefficient is: [Equation 1] and ρ ＸＹ is the Pearson correlation coefficient, X and Y represent the encoded features and the corresponding phenotype data, respectively. cov(*) represents the covariance, σ represents the standard deviation, The plant phenotype prediction method according to claim 3 .

5. Step (2) is (2.1) For each omics feature obtained in step (1), [Equation 2] performing a normalization process based on f ｉｊ represents the j-th omics feature of the i-th sample, where i = 1...M, M represents the number of samples, and j = 1...N, N represents the number of omics; f ｉｊ ′ represents the j-th omics feature after normalization of the i-th sample, u ｊ is the average value of all samples of the jth omics, var ｊ is the standard deviation of all samples of the j-th omics, and (2.2) For each sample, calculating a cosine similarity between the normalized omics features calculated in step (2.1) as a first similarity, the calculation formula being: [Equation 3] and similarity is the cosine similarity, A and B represent the omics features after normalization. A and B are the dot product, where ||A|| represents the norm corresponding to the omics feature A after normalization; (2.3) for each sample, constructing graph structure data of each sample, where different omics of the sample are graph nodes, and the first similarity between the different omics features obtained in step (2.2) is the weight of the edge of the graph; The plant phenotype prediction method according to claim 1 .

6. The first loss includes a predicted regression loss and a Pearson correlation coefficient loss, and its formula is: [Equation 4] and Loss1 represents the first loss, M is the number of samples, Y ｉ is the true phenotype data of node sample i, Y ｉ ' is the first phenotype prediction value of node sample i, [Equation 5] is the mean value of the true phenotypic data of all samples, [Equation 6] is the average of the first phenotype prediction values ​​of all samples, α is the adjustment factor between the predicted regression loss and the Pearson correlation coefficient loss, The plant phenotype prediction method according to claim 1 .

7. Step (5) is (5.1) A step of generating a global identifier embedding feature by initializing using random normal distribution parameters, where the values ​​of each dimension of the global identifier embedding feature are all taken from random values ​​generated by normal distribution, and the feature length is the same as the feature length of each node extracted in step (3); (5.2) for each sample, calculate a first attention matrix of the sample based on the features of all nodes of the sample extracted in step (3); (5.3) for each sample, the feature of each node of the sample extracted in step (3) is set as a local identifier embedding feature, the global identifier embedding feature obtained in step (5.1) and the local identifier embedding feature of each node are input into a Transformer network including a self-attention module and a feed-forward neural network module, the self-attention module calculates a second attention matrix corresponding to the sample, the second attention matrix and the first attention matrix obtained in step (5.2) are combined to update the global identifier embedding feature and the local identifier embedding feature of each node, and the feed-forward neural network module obtains the updated global identifier embedding feature and the local identifier embedding feature of each node; (5.4) For each sample, calculate a cosine similarity between the updated global identifier embedding feature obtained in step (5.3) and the local identifier embedding feature of each node as a second similarity; combine the local identifier embedding feature and the global identifier embedding feature of the two nodes with the highest second similarity to obtain a second combined feature; input the second combined feature into two second fully connected layers to obtain a second phenotype prediction value; and calculate a second loss based on the second phenotype prediction value and corresponding true phenotype data. The plant phenotype prediction method according to claim 1 .

8. The second loss is the predicted regression loss, whose formula is: [Equation 7] and Loss2 is the second loss, M is the number of samples, Y ｉ is the true phenotype data of node sample i, Y ｉ '' is the second phenotype predicted value of node sample i, The plant phenotype prediction method according to claim 1 .

9. 1. A plant phenotype prediction device comprising: one or more processors; and a memory connected to the processors, the memory is used to store program data; The processor is used to execute the program data to implement the plant phenotype prediction method according to any one of claims 1 to 8. A plant phenotype prediction device characterized by:

10. A computer-readable storage medium storing a program, the program being executed by a processor to perform the plant phenotype prediction method according to any one of claims 1 to 8. A computer-readable storage medium comprising:

Images