A genomic prediction method and model integrating maize genotype and environmental cross-modal features

Abstract

The present invention provides a genomic prediction method and model that integrates cross-modal features of corn genotypes and environments. It uses deep learning methods to model the complex nonlinear transformation relationship between different omics data, including genome and environment groups. By constructing a GECF_GP model to integrate the characteristics of corn cross-genotype and environment interactions, it better explores temporal characteristics and important growth periods related to phenotypes, and realizes the function of predicting the phenotypes of complex agronomic traits of corn. The present invention demonstrates strong robustness in different experimental settings, has superiority in genomic prediction of G×E interactions, explores important growth periods related to phenotypes and the performance of predicting phenotypes across different years, and provides new ideas for genomic prediction and analysis methods. It further provides theoretical and application value references for the adaptability analysis of corn in different environments and the exploration of key growth periods, accelerating the development of the field of smart breeding.

Description

Technical Field

[0001] The present invention belongs to the technical field of genome prediction, and specifically relates to a genome prediction method and model that integrates corn genotype and environmental cross-modal features. Background Art

[0002] As one of the world's most important food crops, maize plays an irreplaceable role in addressing food security and meeting the world's food needs. However, with the intensification of global climate change, food production faces unprecedented challenges. Extreme weather and climate instability increase the uncertainty of food production, posing a major threat to global food security. Therefore, genomic prediction methods for genotype-by-environment (G×E) interactions are of great significance for their application in maize breeding and production.

[0003] Currently, genomic prediction methods for G×E interactions primarily rely on classical statistical methods. These methods can effectively address linear mechanisms within G×E interactions, but they are unable to delve into the complex nonlinear mechanisms underlying these interactions. Existing G×E genomic prediction methods primarily extract features from environmental data using full-growing-period EC covariates or sub-period EC covariates. These methods are unable to capture the complex temporal transformations within environmental data. Consequently, existing methods have limited exploration of the optimal growth period for phenotypes, and prediction accuracy needs to be further improved. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a genomic prediction method and model that integrates corn genotype and environmental cross-modal characteristics, so as to predict the phenotype of complex agronomic traits of corn.

[0005] The technical solution adopted by the present invention to solve the above technical problems is: a genome prediction method that integrates corn genotype and environmental cross-modal features, comprising the following steps:

[0006] S1: Sequencing maize to obtain whole-genome SNP data, and collecting environmental data and field agronomic traits of maize planting sites throughout the entire growth period;

[0007] S2: Extract and fuse multidimensional features between sites of whole genome SNP data from different dimensions including height, width and channel to obtain genotype features;

[0008] S3: Capture and integrate the temporal characteristics of multiple environmental factors of the environmental data throughout the entire growth period to obtain temporal environmental characteristics. The specific steps are as follows:

[0009] S31: K-means clustering method was used to group similar environmental factors into the same category;

[0010] S32: Capturing the full growth period temporal characteristics of each environmental factor within the cluster through the autocorrelation mechanism;

[0011] S33: Fusion of temporal features of multiple environmental factors within a cluster through a fully connected neural network;

[0012] S34: Fusion of inter-class environmental factor features through the cross-attention mechanism;

[0013] S4: Fusion of the genotype features obtained in step S2 and the temporal environmental features obtained in step S3 through the Concat mechanism;

[0014] S5: The complex relationships between different omics features are integrated through a multi-layer fully connected neural network; the phenotypic values ​​of specific traits of specific maize materials are predicted and output through the last layer of the fully connected neural network.

[0015] According to the above scheme, in step S1, the specific steps are:

[0016] S11: Whole-genome SNP data obtained by sequencing the maize population; random screening of the whole-genome SNP data to obtain genotype data G of L SNP sites in the whole genome L ={g1,g2…g L}, g i ∈{0,1,2}, i∈{1,2…L}, the density of the SNP site data after screening is consistent with the density distribution of the data before screening; the SNP site data after screening is converted into a matrix X∈R with a length of L and a number of channels C L×C , R is the field of real numbers;

[0017] S12: Collect environmental factors related to light, temperature, water, air and soil during the whole growth period of the corn planting site; let the i-th environmental factor on the Mth day be x M , the i-th environmental factor within M days is represented as an array f i ={x1,x2…x M}, then the data of f environmental factors in M ​​days is a two-dimensional matrix E∈R f×M ;

[0018] S13: Collect phenotypic values ​​of field agronomic traits including plant height and yield at maize planting sites.

[0019] Furthermore, in step S2, the specific steps are:

[0020] S21: Matrix X∈R constructed for whole genome SNP data L×C Perform Patch operation and divide the blocks into small blocks X (1) ∈R H×C×P, L = H × P, each small block contains P SNP data, and the height of each small block is H;

[0021] Let W embedding is the weight of the Embedding operation, b embedding is the bias of the Embedding operation, D is the dimension after Embedding; for X (1) Through the Embedding operation, the high-dimensional features are mapped to obtain the encoded block embedding feature Y embedding ∈R H×C×D :

[0022] Y embedding =W embedding X (1) +b embedding (1);

[0023] S22: extracting multi-dimensional intra-block features including height, width and channel from the encoded block embedding feature Y, capturing the position information of the block and the feature relationship between the blocks;

[0024] S23: Perform a global average pooling operation on the feature vector Y extracted in step S22, and set Y c,h,d is the element of the cth channel, hth row, and dth column of the input feature vector, then the genotype feature G output by global feature fusion is:

[0025]

[0026] Furthermore, in step S22, the specific steps are:

[0027] S221: Let x i represents the input of the neurons in the i-th layer, σ i 2 Represents the mean and variance of the input, and ε takes the default value of 1e-5; LayerNorm is performed on the encoded block embedding features to speed up model training and keep the translation and scaling of features unchanged:

[0028]

[0029] S222: Perform channel dimension transformation on the features after layer normalization through Forward Transpose operation, transpose the corresponding dimension to the end, and output it as X (2) ∈R H×D×C ;

[0030] S223: Let W1 and W2 represent the weights of the two-layer linear mapping, b1 and b2 are the biases of the two-layer linear mapping, σ represents the activation function Gelu, and the features after channel dimension transformation are subjected to two-layer channel dimension MLP operation to extract the important features X of the channel dimension. (3) for:

[0031] X (3) =σ(W2σ(W1X (2) +b1)+b2) (3),

[0032] Among them, the activation function Gelu is used to nonlinearize the captured linear features:

[0033] Gelu(x)=xP(X≤x)X~N(0,1) (4);

[0034] S224: Restore the channel dimension through the Back transpose operation and convert the X output of the MLP operation (3) ∈R H×D×C Adjust to Y (1) ∈R H×C×D As input to the next layer;

[0035] S225: Repeat steps S221 to S224 to extract features of different dimensions including height, width and channel in sequence.

[0036] According to the above solution, in step S31, the specific steps are:

[0037] S311: Let f j is the jth environmental factor, μ j is the center point of the jth class, and the index c of the cluster to which the jth data point belongs j for:

[0038]

[0039] C i is the set of environmental factors of type i, randomized class center μ i for:

[0040]

[0041] S312: Let k be the number of categories, assign the environmental factor to the nearest category, and find the objective function J:

[0042]

[0043] S313: define a new cluster center as the center of the environmental factors belonging to the cluster;

[0044] S314: Repeat steps S312 and S313 to minimize the objective function J.

[0045] Furthermore, in step S32, the specific steps are:

[0046] The environmental factor sequence f is transformed into i The linear mapping is query vector Q, key vector K, value vector V, let W q is the weight matrix of the query vector Q, W k is the weight matrix of the key vector K, W v is the weight matrix of the value vector V:

[0047] Q=W q f i (9),

[0048] K=W k f i (10),

[0049] V=W v f i (11);

[0050] Based on the Wiener-Khinchi theory, the time domain is converted to the frequency domain through the fast Fourier transform FFT, and the autocorrelation of the time series is calculated in the frequency domain. * is the conjugate transform of FFT, j is the imaginary unit, n is the index of the median of the sequence, m is the index of the discrete frequency, and S is the intensity of the different frequency components in the sequence. Q,K for:

[0051]

[0052] The frequency domain is transformed into the time domain by inverse fast Fourier transform IFFT, and the autocorrelation function R of the time series is obtained. Q,K , the autocorrelation function describes the similarity of the sequence at different time delays, that is, the confidence of the time delay in the sequence:

[0053]

[0054] Select the l time delays with the highest confidence level and take

[0055] τ1,τ2,..,τ l =argTopk(R Q,K (τ)) (14),

[0056] The similarity value of the delay is normalized by the softmax function to obtain the weight of the delay:

[0057]

[0058] V(t-τ i ) is V at τ i After the delay, the rolling sequence is weighted and summed, and the features of the subsequences with different delays are aggregated to obtain the final result of feature extraction, Auto-Correlation:

[0059]

[0060] Furthermore, in step S34, the specific steps are:

[0061] Learn the correlation between environmental factors and perform weighted fusion of features of different categories based on the learned weights to extract inter-class feature information;

[0062] Perform a cross-attention mechanism on k categories to extract features and obtain k*(k-1) feature vectors e:

[0063] CrossAttention(Q,K,V)=SoftMax(Q·K T )·V (17),

[0064] e i =CrossAttention(Q i ,K i+1 ,V i+1 ),i∈{1,2,...,n} (18), and fuse these feature vectors through the Concat layer to obtain the temporal environment feature E:

[0065] E=Concat(e1,e2,…,e k*(k-1) ) (19).

[0066] Furthermore, in step S4, the specific steps are:

[0067] The Concat mechanism is used to fuse the genotype feature G extracted in step S2 and the temporal environmental feature E extracted in step S3:

[0068] F = Concat(G, E) (20).

[0069] Furthermore, in step S5, the specific steps are:

[0070] Let W MLP1 、W MLP2 and W MLP3 are the weights of the three-layer fully connected neural network, b MLP1 、b MLP2 and b MLP3are the biases of the three-layer fully connected neural network respectively. The fusion feature F obtained in step S4 is fused with the complex relationship between different omics features through a multi-layer fully connected neural network. The phenotypic value Phe of the specific trait of the specific corn material is predicted and output by the last layer of the fully connected neural network:

[0071] Phe=W MLP3 (W MLP2 BatchNorm(W MLP1 F+b MLP1 )+b MLP2 )+b MLP3 (twenty one);

[0072] Let mean[x], var[x] be the global mean and variance obtained during network training, and the default value of ε is 1e-05; γ is a scaling factor that allows the model to recover some information lost during normalization; β is a translation factor that allows the model to adjust the distribution of inputs; dropout and batch normalization BatchNorm operations are used to prevent gradient vanishing and overfitting while capturing nonlinear features:

[0073]

[0074] Let N represent the number of materials involved in the MSELoss calculation, and use the mean of the sum of squares of the differences between the predicted phenotypic value Phe and the true phenotypic value y to measure the network loss. The mean square error loss function MSELoss is:

[0075]

[0076] A genomic prediction model that integrates maize genotype and environmental cross-modal features is constructed based on a genomic prediction method that integrates maize genotype and environmental cross-modal features;

[0077] It includes genotype feature extraction module, temporal environmental feature extraction module and cross-omics feature fusion module;

[0078] The genotype feature extraction module is a fully connected neural network module (MDMLP) for multidimensional spatial feature fusion, including a Patch Embedding layer, N MDblock layers, a global average pooling layer, and a fully connected layer, which is used to extract and fuse the multidimensional features of SNP site data. The Patch Embedding layer includes a Patch layer and an Embedding layer; the Patch layer is used for block division; the Embedding layer is used for encoding; each MDblock layer includes a token mixing layer (Base-dim MDLayer), a height mixing layer (Height MDlayer), and a channel mixing layer (Channel MDlayer), which are used to extract input features in different dimensions respectively; the global average pooling layer is used for global feature fusion, improving the model's ability to represent global SNP features and reducing the impact of position bias on feature extraction.

[0079] The temporal environmental feature extraction module first uses the K-means clustering method to cluster environmental factors into k categories. Secondly, it captures the temporal features of each environmental factor within the category throughout the growth period through the autocorrelation mechanism. Then, it fuses the full growth period features of multiple environmental factors within the category through a fully connected neural network. Finally, it fuses the features of environmental factors between categories through the cross-attention mechanism, thereby capturing the temporal features of multiple environmental factors throughout the growth period.

[0080] The cross-omics feature fusion module regards each omics as a modality and uses the Concat mechanism to construct a cross-omics feature fusion layer to effectively fuse the captured genotypes with environmental temporal characteristics and output phenotypic values ​​for predicting specific traits of corn.

[0081] The beneficial effects of the present invention are:

[0082] 1. The present invention provides a genomic prediction method and model for fusing cross-modal features of corn genotypes and environments. Deep learning methods are used to model the complex nonlinear transformation relationships between different omics data including genome and environmental groups. By constructing a GECF_GP model to integrate the features of corn cross-genotype and environmental interactions, the temporal characteristics and important growth periods related to phenotypes are better mined, realizing the function of predicting the phenotypes of complex agronomic traits of corn.

[0083] 2. The present invention demonstrates strong robustness across diverse experimental settings. Based on the maize976 corn population dataset, the patterns of untested genotypes in observed environments and tested genotypes in unobserved environments demonstrate that the present invention outperforms six existing cutting-edge genomic prediction methods in predicting field agronomic trait phenotypes. It also demonstrates superiority in genomic prediction of G×E interactions, providing new insights into genomic prediction analysis methods and a theoretical basis for analyzing corn adaptability in diverse environments.

[0084] 3. This invention explores the key growth periods associated with phenotypes and the performance of predicting phenotypes across different years. Regarding key growth periods, the present invention shows significant differences in the predicted phenotypic distributions for the entire growth period and six sub-growth periods, with the prediction accuracy for sub-growth periods slightly higher than that for the entire growth period. Furthermore, the present invention's model, trained on specific genotypes in an unobserved environment, is more suitable for phenotypic prediction across years than the model trained on untested genotypes in an observed environment. This provides valuable application reference for exploring key growth periods for corn and accelerates the development of intelligent breeding. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 is a flow chart of an embodiment of the present invention.

[0086] Figure 2 It is a hierarchical framework diagram of the GECF_GP model of an embodiment of the present invention.

[0087] Figure 3 3 is a graph comparing the performance of the embodiments of the present invention and different existing methods for predicting the Maize976 population in an observed environment using untested genotypes.

[0088] Figure 4 3 is a graph comparing the performance of the embodiment of the present invention and different existing methods for predicting the Maize976 population in an unobserved environment using a test genotype.

[0089] Figure 5 1 is a performance evaluation graph of the embodiment of the present invention for predicting the phenotype of the Maize976 population during the entire growth period and six sub-growth periods.

[0090] Figure 6 This is a performance evaluation diagram of a model trained with untested genotypes in an observed environment and a model trained with known genotypes in an unobserved environment in predicting phenotypes across years according to an embodiment of the present invention. DETAILED DESCRIPTION

[0091] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0092] Example 1

[0093] See also Figure 1 , this embodiment includes the following steps:

[0094] S1: Sequencing to obtain single nucleotide variant (SNP) data of the whole maize genome, collecting environmental data of each planting site throughout the growth period and field agronomic traits of each planting site;

[0095] The single nucleotide variation data (SNP) of the whole genome of the corn population obtained by sequencing was randomly screened to obtain the genotype data G of L SNP sites in the whole genome. L ={g1,g2…g L}, g i ∈{0,1,2}, i∈{1,2…L}, the density of SNP sites after screening is consistent with the density distribution before screening; convert it into a matrix X∈R with a length of L and a number of channels C L×C , R is the field of real numbers;

[0096] For environmental data, collect the environmental factors related to light, temperature, water, air and soil every day during the corn planting period; let the i-th environmental factor on the Mth day be x M , the i-th environmental factor within M days is represented as an array f i ={x1,x2…x M}, then the data of f environmental factors in M ​​days is a two-dimensional matrix E∈R f×M .

[0097] For the field agronomic trait data of each breeding material, the phenotypic values ​​of agronomic traits such as plant height and yield were collected at each planting location.

[0098] S2: In the GECF_GP model, for the input whole-genome SNP data, the multidimensional spatial feature fusion fully connected neural network (MDMLP) module is used to effectively extract the complex relationship features between SNP sites from different dimensions (height, width and channel);

[0099] MDMLP includes Patch Embedding, MDblock, global average pooling, and fully connected layers to extract and fuse multidimensional features of SNP locus data. The specific implementation steps are as follows:

[0100] S201: Matrix X∈R constructed from SNP data L×C Input to the Patch Embedding layer, use Patch to divide X into blocks to obtain X (1) ∈R H×C×P , where L = H × P; each block contains P SNP data, and the height of each block is H. (1) The high-dimensional features are mapped through the Embedding layer to obtain the encoded block embedding feature Y embedding , the specific implementation process is shown in formula (1).

[0101] Y embedding =W embedding X (1) +b embedding (1)

[0102] In formula (1), W embedding is the weight of the embedding layer, b embedding is the bias of the embedding layer, D is the dimension after embedding, Y embedding ∈R H×C×D .

[0103] S202: The encoded block embedding features are input into the MDblock layer for feature extraction. Each MDblock is composed of three MDLayers. These three MDLayers represent the input features and extract features at different dimensions, including the token mixing layer (Base-dim MDLayer), the height mixing layer (Height MDLayer), and the channel mixing layer (Channel MDLayer). By applying MLP on multiple dimensions, not only can intra-block features be captured, but also the block location information and the feature relationships between blocks can be captured.

[0104] First, the feature vector of the input MDLayer is subjected to layer normalization (LayerNorm) to speed up the model training and keep the translation and scaling of the features unchanged. The implementation process is shown in formula (2).

[0105]

[0106] In formula (2), x i represents the input of the neurons in the i-th layer, σ i 2 Represents the mean and variance of the input, and ε takes the default value 1e-5.

[0107] Secondly, for the features after layer normalization, the input is dimensionally changed to facilitate operations on the corresponding dimensions. Specifically, the input data is represented by Y∈R H×C×D After the output of the Embedding layer and layer normalization, the data will be re-arranged by the Forward Transpose layer, and the corresponding dimension will be transposed to the end. H×C×D , the output is X (2) ∈R H×D×C .

[0108] Furthermore, the features after channel dimension transformation are subjected to two-layer channel dimension MLP operation to extract important features of the channel dimension. The specific implementation process is shown in formula (3).

[0109] X (3) =σ(W2σ(W1X (2) +b1)+b2) (3)

[0110] In formula (3), W1 and W2 represent the weights of the two-layer linear mapping, b1 and b2 are the biases of the two-layer linear mapping, and σ represents the activation function Gelu. This operation can nonlinearize the captured linear features. The specific implementation process is shown in formula (4).

[0111] Gelu(x)=xP(X≤x) X~N(0,1) (4)

[0112] Finally, the channel dimension is restored by back transpose, and the output X of MLP is (3) ∈R H×D×C Adjust to Y (1) ∈R H×C×D as input to the next layer.

[0113] S203: Repeat the above S202 process to extract the multi-dimensional features of the token mixing layer, the height mixing layer and the channel mixing layer in turn.

[0114] S204: Input the feature vector Y extracted by the MDBlock layer into the global average pooling layer to achieve global feature fusion, improve the model's ability to represent global SNP features, and reduce the impact of position bias on feature extraction. The specific implementation process is shown in formula (5).

[0115]

[0116] In formula (5), Y c,h,d represents the element in the cth channel, hth row, and dth column of the input vector, and G represents the output feature after global average pooling, that is, the extracted SNP feature.

[0117] S3: For the input full-growing-period environmental data, the environmental factors are first clustered into k categories using the K-means clustering method. Secondly, the full-growing-period temporal characteristics of each environmental factor within the class are captured through the autocorrelation mechanism. The full-growing-period characteristics of multiple environmental factors within the class are then fused through a fully connected neural network. Finally, the cross-attention mechanism is used to fuse the environmental factor characteristics between classes, thereby capturing the temporal characteristics of multiple environmental factors throughout the whole growing period.

[0118] S301: Cluster the environmental factors into k categories using the K-means clustering method. Similar environmental factors are grouped into the same category to facilitate further analysis and processing.

[0119] The K-means method consists of the following steps:

[0120] First, randomize the cluster center μ i, then assign the environmental factor to the nearest class, and finally define the new class center as the center of the environmental factor belonging to this class, and repeat steps 2 and 3 to minimize the objective function J. The formal definitions of the above steps are shown in formulas (6), (7), and (8).

[0121]

[0122] In formula (6), f j Refers to the jth environmental factor, μ j represents the center point of the jth class, where c j Indicates the index of the cluster to which the jth data point belongs. In formula (7), C i represents the set of environmental factors of the i-th category, and k is the number of categories in formula (8).

[0123] S302: The full-growing period time series characteristics of each environmental factor in the class are captured through the autocorrelation mechanism. The autocorrelation mechanism is a further optimization of the shortcomings of the self-attention mechanism in processing time series data. The mechanism calculates the similarity between sequences based on the relationship of periodic dependency and aggregates similar subsequences through time delay information. The autocorrelation mechanism uses its inherent sparsity and subsequence-level similarity for characterization, thereby improving computational efficiency and information utilization. The autocorrelation mechanism converts the environmental factor sequence f i Perform linear mapping to query vector Q, key vector K, value vector V, let W q is the weight matrix of the query vector Q, W k is the weight matrix of the key vector K, W v is the weight matrix of the value vector V:

[0124] Q=W q f i (9)

[0125] K=W k f i (10)

[0126] V=W v f i (11)

[0127] Based on the Wiener-Khinchi theory, the autocorrelation of a time series can be efficiently calculated in the frequency domain by converting the time domain to the frequency domain through the fast Fourier transform (FFT). The autocorrelation function of the time series can be obtained by converting the frequency domain to the time domain through the inverse Fourier transform (IFFT). The autocorrelation function describes the similarity of the sequence at different time delays, that is, the confidence level of the time delay in the sequence. The specific implementation process is shown in formulas (12) and (13).

[0128]

[0129] In formula (12), FFT refers to fast Fourier transform, FFT * represents its conjugate transformation, S Q,K represents the intensity of different frequency components in the sequence, j is the imaginary unit, n is the index of the sequence median, m is the index of the discrete frequency, and in formula (13) IFFT refers to the inverse fast Fourier transform, R Q,K is the similarity of the sequence to itself at different time delays.

[0130] Select the l time delays with the highest confidence, normalize their similarity values ​​through the softmax function to obtain the weight of the time delay, further weight the sum of similar subsequences corresponding to the time delay, and then aggregate the features of subsequences with different time delays to obtain the final result Auto-Correlation after feature extraction. Therefore, the time complexity of the autocorrelation mechanism is O(DlogD), which is lower than that of the self-attention mechanism O(D 2 ) The time complexity has been greatly improved. The specific process is shown in formula (14), formula (15) and formula (16).

[0131] τ1,τ2,..,τ l =argTopk(R Q,K (τ)) (14)

[0132]

[0133] In formula (16), V(t-τ i ) indicates that V is in τ i Delayed scrolling sequence.

[0134] S303: using a fully connected neural network to fuse the temporal characteristics of multiple environmental factors extracted within the class throughout the growth period;

[0135] S304: A cross-attention mechanism is used to fuse inter-class environmental factor features, thereby capturing the temporal characteristics of multiple environmental factors throughout the entire growth period. This mechanism aims to learn the correlations between environmental factor classes and perform a weighted fusion of features from different categories based on the learned weights, thereby extracting inter-class feature information to enhance clustering effectiveness and feature expression capabilities.

[0136] Specifically, this method maps the feature vector of one category to vector Q through a linear layer, and maps the feature vector of the other category to vectors K and V. The specific implementation process is shown in formulas (9), (10), and (11). The specific process of cross-attention is shown in formula (17).

[0137] CrossAttention(Q,K,V)=SoftMax(Q·K T )·V (17)

[0138] For k categories, after extracting features through the cross-attention mechanism between each two, k*(k-1) feature vectors e can be obtained. Subsequently, these feature vectors are fused through the Concat layer to obtain the final environment feature vector E. The specific implementation process is shown in Formulas (18) and (19).

[0139] e i =CrossAttention(Q i ,K i+1 ,V i+1 ),i∈{1,2,...,n} (18)

[0140] E=Concat(e1,e2,…,e k*(k-1) ) (19)

[0141] S4: Concat mechanism is used to construct a cross-omics feature fusion layer to effectively integrate the captured genotype features with temporal environmental features;

[0142] Based on the genotype feature G extracted in step S2 and the temporal environmental feature E fused with multiple environmental factors extracted in step S3, the Concat mechanism is used to achieve the fusion of cross-genomic and environmental omics features. The specific implementation is shown in formula (20).

[0143] F=Concat(G,E) (20)

[0144] S5: The output layer uses a multi-layer fully connected neural network to further integrate the complex relationships between different omics features. The last layer of the fully connected neural network is used to predict the phenotypic value of a specific maize material.

[0145] Let W MLP1 、W MLP2 and W MLP3 are the weights of the three-layer fully connected neural network, b MLP1 、b MLP2 and b MLP3 The fusion feature F obtained in step S4 is further integrated through a multi-layer fully connected neural network to enhance the complex relationships between different omics features. Furthermore, the final fully connected neural network layer is used to predict the phenotypic value Phe for a specific trait of a specific maize material. The specific implementation is shown in Equation (21).

[0146] Phe=W MLP3 (W MLP2 BatchNorm(WMLP1 F+b MLP1 )+b MLP2 )+b MLP3 (twenty one)

[0147] During network training, to prevent gradient vanishing and overfitting while capturing nonlinear features, dropout and batch normalization (BatchNorm) operations are used in the network. The batch normalization implementation process is shown in formula (22). Dropout is another commonly used regularization technique that randomly sets the outputs of some neurons in the network to zero to reduce the dependencies between neurons and prevent overfitting.

[0148]

[0149] In formula (22), mean[x], var[x] are the global mean and variance obtained during network training; the default value of ε is 1e-05; γ is a scaling factor that allows the model to recover some information lost during normalization; β is a translation factor that allows the model to adjust the distribution of the input.

[0150] When calculating the network loss, the mean of the sum of squares of the differences between the predicted phenotypic value Phe and the true phenotypic value y of the material is used for measurement, that is, the mean square error loss function MSELoss, and the specific implementation is shown in formula (23).

[0151]

[0152] In formula (23), N represents the number of materials involved in the MSELoss calculation.

[0153] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0154] Example 2

[0155] See also Figure 2 In this example, the genomic prediction model GECF_GP (genotype environment cross-modal feature fusion genetic prediction) of maize genotype and environment cross-modal feature fusion is constructed through the steps of Example 1, including a genotype feature extraction module, a temporal environmental feature extraction module, and a cross-omics (each omics is regarded as a modality) feature fusion module.

[0156] The genotype feature extraction module effectively extracts the complex relationship features between SNP sites from different dimensions (height, width and channel) through the multidimensional space feature fusion fully connected neural network (MDMLP) module.

[0157] The temporal environmental feature extraction module first uses the K-means clustering method to cluster environmental factors into k categories. It then uses an autocorrelation mechanism to capture the temporal features of each environmental factor within a category over the entire growth period. A fully connected neural network is then used to fuse the full growth period features of multiple environmental factors within a category. Finally, a cross-attention mechanism is used to fuse the features of environmental factors across categories, thereby capturing the temporal features of multiple environmental factors over the entire growth period.

[0158] The cross-omics feature fusion module uses the Concat mechanism to construct a cross-omics feature fusion layer to effectively fuse the captured genotypes with environmental temporal features.

[0159] Example 3

[0160] This example, based on the steps of Example 1, performs genome prediction on the genotype data of the Maize976 corn population, environmental data from 11 environments planted in 2011 and 2012, and field agronomic trait data. The prediction results are evaluated as follows:

[0161] The first study evaluated the phenotypic prediction performance of untested genotypes in an observed environment. The phenotypic prediction accuracy of GECF_GP was compared with six cutting-edge G×E genomic prediction methods in the Maize976 population. Figure 3 The six G×E genomic prediction methods compared included the main effect EAD plus GE bias model based on the best linear unbiased prediction of the genome (EAD+GE(GB)), the main effect EAW plus GE reaction norm model based on the best linear unbiased prediction of the genome (EADW+GW(GB)), the main effect EAD plus GE bias model based on the Gaussian kernel (EAD+GE(GK)), the main effect EAW plus GE reaction norm model based on the Gaussian kernel (EADW+GW(GK), the main effect EAD plus GE bias model based on the deep kernel (EAD+GE(DK), and the main effect EAW plus GE reaction norm model based on the deep kernel (EADW+GW(DK)).

[0162] For plant height (PH), GECF_GP's prediction accuracy ranged from 0.64 to 0.87 across 11 environments, while the accuracy of the other six statistical methods ranged from 0.489 to 0.838. All methods had lower prediction accuracy in GX2011, HB2011, and SC2011, but higher accuracy in DHN2011, CQ2011, and HN2011. GECF_GP outperformed all compared methods by at least 3% in each environment. Overall, GECF_GP's average accuracy was 0.7744, outperforming all compared methods. For yield (EW), GECF_GP's prediction accuracy ranged from 0.296 to 0.5154 across seven environments, while the accuracy of the other six statistical methods ranged from 0.113 to 0.481. Across all environments, DHN2011 and YN2012 showed high prediction accuracy, while CQ2012, HN2012, and YN2011 showed poorer prediction accuracy. For EW trait prediction, GECF_GP also achieved higher accuracy than all compared methods in each environment. Among the compared methods, EAD+GE(DK) still had the highest average accuracy, at 0.3915, slightly lower than GECF_GP. EADW+GW(GB) had the poorest prediction accuracy, with an average accuracy of 0.3317, significantly lower than GECF_GP.

[0163] The second evaluation tested the phenotype prediction performance of genotypes in unobserved environments. The prediction accuracy of GECF_GP was compared with six cutting-edge G×E genomic prediction methods in the Maize976 population (e.g. Figure 4 shown).

[0164] For pH prediction, GECF_GP demonstrated the best performance. In the 2011 and 2012 environments, GECF_GP's average prediction accuracy improved by 1% and 0.6%, respectively, compared to the best comparison method, slightly exceeding all comparison methods. Although GECF_GP's accuracy was 3% lower than EAD+GE(DK) in the 2011 GX2011 environment and slightly lower than EADW+GW(GB) (0.874) in the 2012 HN2012 environment (0.873), GECF_GP's overall average accuracy remained higher than all comparison methods. For EW prediction, all six leading G×E genomic prediction methods exhibited significant instability, with negative correlations observed across several environments. For example, in the 2011 DHN environment, EADW+GW(GK) achieved a prediction accuracy of -0.3219. Although GECF_GP only achieved higher prediction accuracy than the other comparison methods in a few environments (such as HN2011 and HN2012), its average accuracy in the 2011 and 2012 environments was 0.458 and 0.425, respectively, representing significant improvements of 12.2% and 7.99% over the best comparison methods (EAD+GE(GK) and EADW+GW(GK)). Furthermore, GECF_GP consistently outperformed the comparison methods in all environments, with no negative correlation observed, demonstrating its stability relative to the comparison methods.

[0165] The third test evaluated the accuracy of GECF_GP in predicting pH under observational conditions for untested genotypes during the full growth period (0-150 days) and six sub-growth periods (0-15 days, 16-35 days, 36-65 days, 66-90 days, 91-120 days, and 121-150 days). The experimental results are shown in Figure 2. Figure 5 shown.

[0166] Across the 11 environments, the accuracy of pH prediction for the full growth period and the six sub-periods was similar across each environment. However, the phenotypic distributions for the full growth period and sub-periods were significantly different (p-value < 0.05). In most environments, the third growth period (36-65 days) was closest to the full growth period prediction, followed by the second growth period (16-35 days) and the fourth growth period (66-90 days). These results are valuable for practical breeding applications, demonstrating that phenotypic prediction using only environmental data from these growth periods can yield results similar to those for the full growth period, thus conserving breeding resources.

[0167] The fourth item evaluates the performance of GECF_GP in predicting pH and EW at the same location in different years.

[0168] In order to evaluate the performance of cross-year prediction, two models were designed for cross-year phenotypic prediction. M1 was a model constructed based on untested genotypes in the observed environment to predict the phenotypic data of 2012 across years, while M2 was a model constructed based on tested genotypes in the untested environment to predict the phenotypic data of 2012 across years. The experimental results are shown in Figure 2. Figure 6 As shown, the prediction accuracy of the M2 evaluation method was significantly higher than that of the M1 evaluation method. In the prediction results for pH traits, the accuracy improved by 4.86%, 14.52%, 4.83%, and 6.68% in the CQ, HB, HN, and YN environments, respectively. In the prediction of EW traits, the accuracy improved by 12.6% and 12.1% in the HN and YN environments, respectively. This method performed better in cross-environment prediction when the environmental dataset included more environmental data. It was found that the model trained on known genotypes in unobserved environments (M2) was more suitable for cross-year phenotypic prediction than the model trained on untested genotypes in observed environments (M1).

[0169] The above embodiments are intended only to illustrate the design concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. The scope of protection of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design concepts disclosed in the present invention are within the scope of protection of the present invention.

Claims

1. A genome prediction method that integrates maize genotype and environmental cross-modal features, characterized by: The following steps are involved: S1: Sequencing maize to obtain whole-genome SNP data, and collecting environmental data and field agronomic traits of maize planting sites throughout the entire growth period; S2: Extract and fuse multidimensional features between sites of whole genome SNP data from different dimensions including height, width and channel to obtain genotype features; S3: Capture and integrate the temporal characteristics of multiple environmental factors of the environmental data throughout the entire growth period to obtain temporal environmental characteristics. The specific steps are as follows: S31: K-means clustering method was used to group similar environmental factors into the same category; S32: Capturing the full growth period temporal characteristics of each environmental factor within the cluster through the autocorrelation mechanism; The specific steps are: Set up the first M Day's i The environmental factors are x M ,Will M Day i Environmental factors are represented as an array f i ={ x 1, x 2… x M }, through the autocorrelation mechanism, the environmental factor sequence f i Linear mapping to query vector Q , key vector K , value vector V ,set up W q is the query vector Q The weight matrix, W k is the key vector K The weight matrix, W v is a value vector V The weight matrix is: (9)， (10)， (11)； Based on Wiener-Khinchi theory, through fast Fourier transform FFT Convert the time domain to the frequency domain and calculate the autocorrelation of the time series in the frequency domain. FFT * for FFT The conjugate transformation of j is the imaginary unit, n is the index of the median value in the sequence, m is the index of the discrete frequency, the intensity of the different frequency components in the sequence S Q,K for: (12)； Through inverse fast Fourier transform IFFT Transform the frequency domain into the time domain to obtain the autocorrelation function of the time series R Q,K , the autocorrelation function describes the similarity of the sequence at different time delays, that is, the confidence of the time delay in the sequence: (13)； Select the one with the highest confidence l Delay, take l =⌊ c × log D⌋: (14)， The similarity value of the delay is passed through softmax Normalize the function to get the weight of the delay: (15)， V ( t-τ i )for V exist τ i After the delay, the rolling sequence is weighted and summed up, and the features of the subsequences with different delays are aggregated to obtain the final result of feature extraction. Auto-Correlation : (16)； S33: Fusion of temporal features of multiple environmental factors within a cluster through a fully connected neural network; S34: Fusion of inter-class environmental factor features through the cross-attention mechanism; S4: Fusion of the genotype features obtained in step S2 and the temporal environmental features obtained in step S3 through the Concat mechanism; S5: The complex relationships between different omics features are integrated through a multi-layer fully connected neural network; the phenotypic values ​​of specific traits of specific maize materials are predicted and output through the last layer of the fully connected neural network.

2. The method for genome prediction based on the fusion of maize genotype and environmental cross-modal features according to claim 1, characterized in that: In the step S1, the specific steps are: S11: Whole genome SNP data obtained by sequencing the maize population; random screening of the whole genome SNP data to obtain the whole genome SNP L Genotype data of SNP loci G L = { g 1, g 2… g L }, g i ∈{0,1,2}, i∈{1, 2…L}, the density of the SNP site data after screening is consistent with the density distribution of the data before screening; the SNP site data after screening is converted to a length of L , the number of channels is C Matrix X ∈ R L×C , R is the field of real numbers; S12: Collect environmental factors related to light, temperature, water, air and soil during the whole growth period of corn planting site; f Environmental factors in M The data within a day is a two-dimensional matrix E ∈ R f×M ; S13: Collect phenotypic values ​​of field agronomic traits including plant height and yield at maize planting sites.

3. The genome prediction method of corn genotype and environmental cross-modal feature fusion according to claim 2, characterized in that: In the step S2, the specific steps are: S21: Matrix constructed for whole genome SNP data X ∈ R L×C Perform Patch operation to obtain small blocks X (1) ∈ R H ×C×P , L=H×P , each block contains P SNP data, the height of each block is H ; set up W embedding is the weight of the Embedding operation, b embedding is the bias of the Embedding operation, and D is the dimension after Embedding; right X (1) Through the Embedding operation, high-dimensional features are mapped to obtain the encoded block embedding features. Y embedding ∈ R H×C×D : (1)； S22: Embed features into the encoded block Y Extract multi-dimensional intra-block features including height, width and channel, capture block location information and feature relationships between blocks; S23: The feature vector extracted in step S22 Y Perform global average pooling operation, set Y c,h,d The first feature vector of the input c Channel, h Row, No. d The elements of the column, the genotype features output by the global feature fusion G for: (5)。 4. The method for genome prediction based on the fusion of maize genotype and environmental cross-modal features according to claim 3, characterized in that: In the step S22, the specific steps are: S221: Set x i Indicates the i The input of the neurons in the layer, σ i 2 represents the mean and variance of the input, ε Take the default value 1e-5; perform layer normalization LayerNorm on the encoded block embedding features to speed up model training and keep the translation and scaling of features unchanged: (2)； S222: Perform channel dimension transformation on the features after layer normalization through Forward Transpose operation, transpose the corresponding dimension to the end, and output X (2) ∈ R H×D×C ; S223: Set and Represent the weights of the two-layer linear mapping, and is the bias of the two-layer linear mapping, Represents the activation function Gelu, which performs two-layer channel dimension MLP operation on the features after channel dimension transformation to extract important features of channel dimension X (3) for: (3)， Among them, the activation function Gelu is used to nonlinearize the captured linear features: (4)； S224: Restore the channel dimension through the Back transpose operation and convert the output of the MLP operation X (3) ∈ R H×D×C Adjust to Y (1) ∈ R H×C×D As input to the next layer; S225: Repeat steps S221 to S224 to extract features of different dimensions including height, width and channel in sequence.

5. The method for genome prediction based on the fusion of maize genotype and environmental cross-modal features according to claim 1, characterized in that: In the step S31, the specific steps are: S311: Set f j For the j environmental factors, μ j For the j The center point of the class, j The index of the cluster to which the data point belongs c j for: (6)； C i For the i A collection of environmental factors, randomized class centers μ i for: (7)； S312: Set k is the number of categories, assign the environmental factor to the nearest class, and find the objective function J : (8)； S313: define a new cluster center as the center of the environmental factors belonging to the cluster; S314: Repeat steps S312 and S313 to minimize the objective function J .

6. The genome prediction method of corn genotype and environmental cross-modal feature fusion according to claim 1, characterized in that: In the step S34, the specific steps are: Learn the correlation between environmental factors and perform weighted fusion of features of different categories based on the learned weights to extract inter-class feature information; right k The cross attention mechanism is used to extract features between the two categories. k *( k -1) eigenvectors e : (17)， (18)， pass Concat Layer fusion of these feature vectors to obtain temporal environment features E : (19) 。 7. The genomic prediction method for maize genotypes by fusing cross-modal features of the environment according to claim 3 or 6, characterized in that: In the step S4, the specific steps are: use Concat Genotype features extracted from the mechanism fusion step S2 G The temporal environment features extracted in step S3 E : (20)。 8. The genome prediction method of corn genotype and environmental cross-modal feature fusion according to claim 7, characterized in that: In the step S5, the specific steps are: set up W MLP1 、 W MLP2 and W MLP3 are the weights of the three-layer fully connected neural network, b MLP1 、 b MLP2 and b MLP3 are the biases of the three-layer fully connected neural network, and the fusion features obtained in step S4 are F The complex relationships between different omics features are integrated through a multi-layer fully connected neural network; the phenotypic values ​​of specific traits of specific maize materials are predicted and output through the last layer of the fully connected neural network. Phe for: (21)； set up mean [ x ] , var [ x ] is the global mean and variance obtained during network training. ε The default value is 1e-05; γ is a scaling factor that allows the model to recover some of the information lost during normalization; β It is a translation factor that allows the model to adjust the distribution of the input; dropout and batch normalization BatchNorm operations are used to prevent gradient disappearance and overfitting while capturing nonlinear features: (22)； set up N Express participation MSELoss Calculate the number of materials using the predicted phenotypic values Phe The true phenotypic value y The mean of the sum of squared differences between them measures the network loss and obtains the mean square error loss function MSELoss for: (23)。 9. A genomic prediction model that integrates maize genotype and environmental cross-modal features, characterized by: A genome prediction method based on the fusion of corn genotype and environmental cross-modal features according to any one of claims 1 to 8; It includes genotype feature extraction module, temporal environmental feature extraction module and cross-omics feature fusion module; The genotype feature extraction module is a fully connected neural network module MDMLP for multidimensional spatial feature fusion, including a PatchEmbedding layer, N MDblock layers, a global average pooling layer, and a fully connected layer, which is used to extract and fuse the multidimensional features of SNP locus data; The Patch Embedding layer includes the Patch layer and the Embedding layer; the Patch layer is used for block division; the Embedding layer is used for encoding; each MDblock layer includes the token mixing layer Base-dim MDLayer, the height mixing layer HeightMDlayer, and the channel mixing layer Channel MDlayer, which are used to extract input features in different dimensions respectively; the global average pooling layer is used for global feature fusion, improving the model's ability to represent global SNP features and reducing the impact of position bias on feature extraction; The temporal environmental feature extraction module first uses the K-means clustering method to cluster environmental factors into k classes. Secondly, the autocorrelation mechanism is used to capture the temporal features of each environmental factor within the class throughout the entire growth period. The full-growth period features of multiple environmental factors within the class are then fused through a fully connected neural network. Finally, the cross-attention mechanism is used to fuse the features of environmental factors between classes, thereby capturing the temporal features of multiple environmental factors throughout the entire growth period. The cross-omics feature fusion module regards each omics as a modality and uses the Concat mechanism to construct a cross-omics feature fusion layer to effectively fuse the captured genotypes with environmental temporal characteristics and output phenotypic values ​​for predicting specific traits of corn.

Images