Deep learning based multi-environmental genome-wide prediction method and system

Abstract

The application discloses a kind of multi-environment whole genome prediction methods based on deep learning, obtain the data set including genotype, phenotype and environment type;Predictive model is constructed, wherein feature extraction module is used to carry out feature extraction to genotype data and environment type data;Feature fusion module is used to fuse genotype and environment type feature;Convolutional neural network module is used to carry out feature extraction and representation learning to fusion feature;Predictive regression module is used to regress the feature that passes through convolutional neural network module, to predict the phenotype value of trait;In a variety of segmentation scenarios, data set is split into training set and verification set, and the phenotype value prediction result of crop trait is obtained after inputting multi-environment whole genome prediction model;Mean square error of prediction result and phenotype value in verification set is calculated as phenotype prediction loss, and the parameters of prediction model are updated by optimizing the loss.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of deep learning and computational biology, specifically to a deep learning-based multi-environment whole-genome prediction method. Background Technology

[0002] Genome-wide selection (GS) breeding is one of the most effective strategies for increasing genetic gain per unit cost. A core step in the GS method is to construct a reasonable genome selection prediction model, in which the interaction between genotype and environment (G×E) plays a crucial role in influencing the accuracy and stability of the genome-wide prediction model. Many methods have been used in single experimental settings, which means that specific prediction models need to be built for different environmental conditions. This significantly limits the effectiveness and versatility of the constructed models in real-world breeding applications.

[0003] Currently, genome-wide selection prediction is typically performed in a single or limited environment, assuming that the environment has no effect on quantitative traits. However, crop phenotypes are determined by their genes, environment, and the interactions between the two. By rationally modeling these interactions, genomic selection models are expected to be able to predict phenotypes more accurately in the development of new varieties or under new environmental conditions.

[0004] In-depth research and analysis have revealed that applying existing deep learning GS methods to crop breeding may encounter the following significant limitations: current deep learning methods fail to design appropriate feature extraction schemes for the characteristics of environmental data structures during modeling, resulting in poor predictive performance; and current genome-wide prediction models typically cannot reveal how or why accurate prediction decisions are made, i.e., they lack interpretability. This black-box characteristic may not be ideal, especially when a deeper understanding of the mechanisms of gene-environment interactions is needed to provide reasonable model conclusions for downstream decision-making. Summary of the Invention

[0005] This invention proposes a multi-environment whole-genome prediction method and system based on deep learning, which solves the problems of poor prediction performance and lack of interpretability of existing prediction models.

[0006] To address the aforementioned technical problems, this invention provides a multi-environment whole-genome prediction method based on deep learning, comprising the following steps:

[0007] Step S1: Obtain a dataset including genotype, phenotype, and environmental type;

[0008] Step S2: Construct a multi-environment whole-genome prediction model including a genotype feature extractor, an environment feature extractor, a feature fusion module, a convolutional neural network module, and a prediction regression module;

[0009] The genotype feature extractor is used to extract features from genotype data;

[0010] The environmental feature extractor extracts environmental information from the entire growth cycle by adding a learnable element to the Transformer encoder and interacting with other environmental elements of variable sequence length.

[0011] The feature fusion module is used to fuse genotype and environmental features;

[0012] The convolutional neural network module is used to extract features and learn representations from the fused features;

[0013] The prediction regression module is used to regress the features passed through the convolutional neural network module in order to predict the phenotypic value of the trait.

[0014] Step S3: Split the dataset into training and validation sets under various segmentation scenarios, and input them into the multi-environment whole genome prediction model to obtain the phenotypic value prediction results of crop traits;

[0015] Step S4: Calculate the mean squared error between the predicted result and the phenotypic values ​​in the validation set as the phenotypic prediction loss, and update the parameters of the prediction model by optimizing this loss.

[0016] Preferably, the genotype feature extractor in step S2 maps each locus of the gene data to a high-dimensional space to extract features from the genotype data:

[0017] ;

[0018] in, For the process Genotype data features after feature transformation; This is genotype data.

[0019] Preferably, the environmental feature extractor in step S2 uses a Transformer encoder, which consists of a multi-head self-attention mechanism and an alternating MLP block layer. By treating environmental data at different times as elements and prepending a learnable element in the environmental data, the environmental information of the entire growth cycle is obtained by interacting with other elements in each layer of the Transformer through the self-attention mechanism.

[0020] Preferably, in each layer of the Transformer encoder, the expression for interacting learnable elements with other elements is:

[0021] ;

[0022] ;

[0023] ;

[0024] ;

[0025] In the formula, For input features; E represents environmental data; For learnable embeddings; For positional embedding; LN is for normalization; For multi-head attention operations; As an intermediate feature; The encoder output; learnable embeddings At the output of the converter encoder, corresponding to MLP is a multilayer perceptron module containing two layers. This is a characteristic of environmental data.

[0026] Preferably, by calculating the self-attention of each element, the self-attention mechanism is obtained by extending it. The expression for calculating the self-attention of each element is as follows:

[0027] ;

[0028] ;

[0029] In the formula, H represents the input dimension, and Q, K, and V represent the query, key, and value, respectively. This is the activation function.

[0030] Preferably, the multiple segmentation scenarios in step S3 include: a normal scenario, a genotype-specific scenario, and an environment-specific scenario; the normal scenario uses five-fold cross-validation to segment all combinations; the genotype-specific scenario performs five-fold segmentation based on the unique identifier of the genotype; and the environment-specific scenario performs five-fold segmentation based on the unique identifier of the environment.

[0031] Preferably, step S4 also employs CNN feature map visualization technology to perform interpretability analysis on the model, providing practical guidance for intelligent planting management.

[0032] The present invention also provides a deep learning-based multi-environment whole genome prediction system, which is applicable to the above-mentioned deep learning-based multi-environment whole genome prediction method. The system includes a feature extraction module, a feature fusion module, a convolutional neural network module, and a prediction regression module.

[0033] The feature extraction module is used to extract features from genotype data and environmental data;

[0034] The feature fusion module is used to fuse genotype and environmental features;

[0035] The convolutional neural network module is used for feature extraction and representation learning of the fused features;

[0036] The prediction regression module is used to regress the features passed through the convolutional neural network module in order to predict the phenotypic value of the trait.

[0037] The present invention also provides an electronic device, comprising: a memory, a processor, and a computer program, wherein the computer program is stored in the memory and configured to be executed by the processor to implement the above-described deep learning-based multi-environment whole-genome prediction method.

[0038] The present invention further provides a computer-readable storage medium storing a computer program, which is executed by a processor to implement the above-described deep learning-based multi-environment whole-genome prediction method.

[0039] The advantages of this invention include at least the following:

[0040] 1. This invention proposes to use Transformer network to extract environmental features. Weather data at different times is treated as a series of elements. When performing location encoding, a learnable element is introduced at the beginning of the sequence. In this way, it interacts with other elements in each layer of Transformer through a self-attention mechanism, and gradually aggregates the information of all elements. In the last layer, this learnable element contains environmental information about the entire growth cycle. This invention uses it as an environmental data feature.

[0041] 2. This invention proposes three test sets and dataset segmentation scenarios, which can more comprehensively examine the performance of the model;

[0042] 3. This invention proposes to use CNN feature map visualization technology to interpret the obtained results and their biological significance from the perspective of observing the regularity of feature patterns extracted by the model, to conduct interpretability analysis of the model, and thus expect to provide accurate practical guidance for intelligent planting management in the future, and enhance the transparency of the model and the credibility of the decision-making process. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0044] Figure 2 This is a diagram of the multimodal deep learning framework used in the embodiments of the present invention;

[0045] Figure 3This is a schematic diagram of a simple method for obtaining average environmental data according to an embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram of the genotype feature extractor according to an embodiment of the present invention;

[0047] Figure 5 This invention provides an environment-type network based on Transformer.

[0048] Figure 6 This is a schematic diagram of the genotype and environmental trait fusion method in an embodiment of the present invention;

[0049] Figure 7 These are schematic diagrams illustrating three training and test set segmentation scenarios proposed in embodiments of the present invention;

[0050] Figure 8 This is a scatter density plot of the model proposed in this embodiment of the invention for predicting yield traits under three scenarios;

[0051] Figure 9 This is a scatter density plot of the model proposed in this embodiment of the invention for predicting plant height traits under three scenarios;

[0052] Figure 10 This is a feature heatmap of the last layer of the convolutional neural network for predicting plant height in an embodiment of the present invention.

[0053] Figure 11 This is a feature heatmap of the last layer of the convolutional neural network for predicting yield traits in an embodiment of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0055] like Figure 1 As shown, this embodiment of the invention provides a multi-environment whole-genome prediction method based on deep learning, including the following steps:

[0056] Step S1: Obtain a dataset including genotype, phenotype, and environmental type.

[0057] Specifically, this embodiment utilizes the spring wheat (Triticum aestivum) database from the International Maize and Wheat Improvement Center (CIMMYT), where genotype and phenotypic data are used for the "Future Innovation Lab" project. The provided data includes biallelic SNPs with a minor allele frequency greater than 1% and a dimension of 18239, encoded in numerical format (0,1,2). CIMMYT also provides AgERA5 platform weather data from 1979 to 2019, including nine climate variables such as rainfall, relative humidity, shortwave radiation, temperature, vapor pressure deficit, and wind speed. The environmental data input for the model is from two weeks before sowing to the end of harvest. After matching the genotype, phenotypic, and environmental data, a total of 58,585 yield records and 78,736 plant height records were obtained.

[0058] Step S2: Construct a multi-environment whole-genome prediction model including a genotype feature extractor, an environmental feature extractor, a feature fusion module, a convolutional neural network module, and a prediction regression module.

[0059] Specifically, to fully utilize the relationship between genotype and environment-based bimodal data, this invention proposes a multimodal deep learning framework. For example... Figure 2 As shown, the framework employing the early fusion strategy includes a genotype feature extractor, an environment feature extractor, a feature fusion module, a convolutional neural network module, and a prediction / regression module. In the feature extraction stage, genotype data features can be obtained separately through the genotype feature extraction network and the environment feature extraction network. and environmental data characteristics ,Will and The fusion is performed to obtain fusion features. The fused features are then fed into a convolutional neural network (CNN), where they are further processed through a series of convolutional layers and activation functions. Finally, using the output features of the CNN, a prediction and regression module predicts the crop's phenotypic values, calculates the mean squared error (MSE) between the predicted and actual phenotypic values ​​as the phenotypic prediction loss, and ultimately updates the model's parameters by optimizing this loss.

[0060] In embodiments of the present invention, a genotype feature extractor such as Figure 3 As shown, genetic data is generally one-dimensional. To represent the number of loci in the gene data, which typically uses a few integer values ​​to represent the variation information at each locus, an embedding layer is used to map each locus of the gene data into a high-dimensional space, specifically as follows:

[0061] ;

[0062] in, For feature transformation The genotypic characteristics after that, After the transformation, the feature height changes from 1 to , For input dimensions.

[0063] Environmental feature extractors such as Figure 4 As shown. Figure 4 Taking a single-layer Transformer encoder as an example, compared to the serial processing of traditional recurrent neural networks (RNNs), the parallel processing of Transformers results in higher training and inference efficiency. Experiments in this invention demonstrate that the Transformer approach improves training speed by more than 20% compared to RNNs. Figure 4 As shown, weather data from different time spans at different times are treated as tokens. This embodiment of the invention proposes adding a learnable token. This token interacts with other tokens at each layer of the Transformer through a self-attention mechanism, gradually aggregating the environmental data across the entire growth cycle. In the final layer, this learnable token contains environmental information about the entire growth cycle. Since the growth cycle of each sample differs, the sequence length of the environmental information is variable, resulting in a non-fixed-length total number of these tokens. This embodiment of the invention uses this as a feature of the environmental data. .

[0064] Specifically, the present invention will embed a learnable embedding Pre-loaded environmental data Among them Sequence length and position embedding representing contextual information Included To retain location information, obtain ,in The Transformer encoder consists of alternating layers of multi-head self-attention and MLP blocks, where the MLP comprises two layers and features GELU nonlinearity. Layer normalization (LN) is applied before each block, and residual connections are applied after each block. Taking a single-layer Transformer as an example, the specific formula for forward inference is as follows:

[0065] ;

[0066] ;

[0067] ;

[0068] ;

[0069] in, ; For input features; E represents environmental data; For learnable embeddings; For positional embedding; LN is for normalization; For multi-head attention operations; As an intermediate feature; The encoder output; learnable embeddings At the output of the converter encoder, corresponding to MLP is a multilayer perceptron module containing two layers. This serves as a representation of the global environmental characteristics during the growth period;

[0070] Self-attention is an important module in the Transformer architecture. It involves the generation of queries, keys, and values, denoted as […]. , and , The generation of all originates from the input. It is essentially a vector generated from input features multiplied by weights. get:

[0071] .

[0072] And self-attention The calculation is using and dot product of vectors Calculated, and according to Scale, and then compare these weights with Multiplying them together yields the self-attention output for each element in the sequence:

[0073] ;

[0074] Based on this mechanism, it is easy to extend to multi-head self-attention mechanisms.

[0075] Compared to VIT, the feature extraction process for environmental data does not go through the Linear Projection of Flattened Patches module. This is because VIT addresses the correlation between images and space, while this invention focuses on mining the temporal relationships between environmental data. The environmental data at each time step is already one-dimensional, and skipping this step makes the model more concise and efficient.

[0076] In this embodiment of the invention, a simple environmental feature extractor scheme is also included, such as... Figure 5 As shown, all weather data within the growth cycle are averaged over time, where the input environment type is considered as follows: ,in The length of the sequence representing environmental information. This refers to nine environmental variables. Consider a simple environmental feature extractor such as... Figure 5 As shown, Averaging over time yields... , Each element in the equation represents the average value of different environmental variables during the growth cycle.

[0077] In this embodiment of the invention, the fusion module is as follows: Figure 6 As shown, to accurately simulate the interaction between genotype and environment, these effects are on genes at the locus level. Specifically, [the following is a list of effects]. Perform a broadcast operation to match Thus obtain Subsequently, regarding and By adding elements one by one, the fusion characteristics are obtained. :

[0078] ;

[0079] Fusion features Input the data into the backbone network for feature extraction and representation learning.

[0080] Step S3: Split the dataset into training and validation sets under various segmentation scenarios, and input them into the multi-environment whole genome prediction model to obtain the phenotypic prediction results of crop traits.

[0081] like Figure 7 As shown, a complete dataset is a combination of genotype and environmental data. To more comprehensively evaluate the model's predictive ability, this invention designs three splitting schemes for the training and test sets to perform five-fold cross-validation analysis. The first case, called the standard scenario, uses five-fold cross-validation to split the data, and the model trained using all the training data from the remaining folds is used to predict the [predicted data]. The first scenario is the data in the validation set. The second scenario, called the gene-specific scenario, involves all records being randomly assigned to five non-overlapping groups based on a unique strain identifier, in this case, germplasm ID. This simulates the problem of predicting the performance of new strains that have not yet undergone any field evaluation, thus enabling the assessment of the model's ability to generalize to new genotypes. The third scenario is the environment-specific scenario, where all records are randomly assigned to five non-overlapping groups based on a unique environment identifier, in this case, location-year. In this method, the environment in the test set does not exist in the training set. This scenario is similar to the prediction problem faced when using current data to predict future strain performance.

[0082] Step S4: Calculate the mean squared error between the predicted result and the phenotypic values ​​in the validation set as the phenotypic prediction loss, and update the parameters of the prediction model by optimizing this loss.

[0083] Meanwhile, to improve the transparency and credibility of the model's decision-making process, this invention employs CNN feature map visualization technology, attempting to explain the obtained results and their biological significance from the perspective of observing the regularity of the feature patterns extracted by the model.

[0084] Figure 8 and Figure 9 The proposed method uses scatter density plots to visualize the predicted and labeled values. These plots effectively reveal variable correlations, distribution density, and central tendency in large-scale datasets through color gradients. The density plots can concisely reflect the captured nonlinear relationships by fitting regression lines. The results demonstrate that the proposed algorithm is more likely to produce high-accuracy trait predictions in regions with dense phenotypic distributions.

[0085] Figure 10 This study reveals that different environmental types lead to different expressions of the same genotype. Each subplot is generated from validation data: six samples from the same genotype and six different environments, visualized by forward reasoning up to the feature map of the last layer of a convolutional neural network. The left and right columns represent low phenotypic values ​​(less than 50) and high phenotypic values ​​(greater than 120) for plant height, respectively. It can be observed that features within the same column exhibit high similarity, while features in the left and right columns show significant differences. This aligns with the common knowledge that the same gene expresses differently in different environments, indicating that the proposed model can respond to different environments and further demonstrating the effectiveness of the instance fusion strategy. Figure 11 As shown, this phenomenon also appears in yield forecasting.

[0086] In this embodiment of the invention, the proposed whole-genome prediction model was compared with other state-of-the-art algorithms. These algorithms are divided into four categories. The first category is the GBLUP (Genomic Best Linear Unbiased Prediction) algorithm based on a mixed linear model, which is a one-step method. The second category is the Bayes method, including algorithms such as BayesRR, BayesB, BayesBpi, and BayesR. The third category is machine learning algorithms, with Support Vector Regression (SVR) and LightGBM selected as representatives. The fourth category of algorithms uses the Multilayer Perceptron (MLP) model. The prediction results of the method in this study for two traits—yield and plant height—in three scenarios are shown in Tables 1 to 6. The prediction accuracy index examined is the Pearson correlation coefficient.

[0087] Table 1. Comparison of the prediction performance of each model for output in a typical scenario.

[0088]

[0089] Table 2 Comparison of yield prediction performance of each model in genotype-specific scenarios

[0090]

[0091] Table 3 Comparison of the predictive performance of each model for output in specific environmental scenarios.

[0092]

[0093] Table 4 Comparison of plant height prediction performance of each model in conventional scenarios

[0094]

[0095] Table 5. Comparison of plant height prediction performance of each model in genotype-specific scenarios.

[0096]

[0097] Table 6. Comparison of plant height prediction performance of each model in specific environmental scenarios.

[0098]

[0099] Overall, deep learning methods show a clear advantage in predictive performance. The methods proposed in this invention, denoted as Ours_Mean and Ours_Transformer, achieve the best performance compared to other methods, reaching a strong correlation of 0.8901. Machine learning methods are in the second tier, with competitive prediction accuracy, but significantly lagging behind deep learning methods. LightGBM, based on an ensemble approach, shows some advantage in SVR. Regression models have the lowest predictive correlation; in the three scenarios, yield and plant height are below 0.3 and 0.5 respectively, which is expected, as linear models like GBLUP and Bayes regression struggle to capture the complex interactions between genotype and environment. It is worth emphasizing that Ours_Transformer significantly improves prediction accuracy compared to Ours_Mean, indicating that the two environment-based networks designed in this invention are effective and can accurately extract features from time-series environmental data throughout the crop growth cycle.

[0100] This invention discloses a multi-environment whole-genome prediction method based on the Transformer framework. Addressing the current shortcomings of deep learning methods in designing corresponding feature extraction schemes for environmental data structures, leading to poor predictive performance, and the lack of interpretability analysis methods for current whole-genome prediction models, this invention utilizes a multimodal deep learning framework for multi-environment whole-genome prediction. The environmental branch network is based on the Transformer model, and its parallel attention mechanism ensures high efficiency in both training and inference phases. This invention also proposes three test and validation set partitioning strategies, each simulating corresponding real-world production scenarios, providing a means for more comprehensive evaluation of multi-environment whole-genome prediction models. This invention uses convolutional neural network feature map visualization technology for model interpretability analysis, interpreting the results and their biological significance by observing the regularity of the extracted feature patterns. The proposed model was compared with several representative algorithms, achieving state-of-the-art strong correlation levels in prediction performance. The results of the proposed interpretability analysis method further confirm the rationality of the model design. This provides important reference for breeders in selecting new varieties for different environments.

[0101] This invention also provides a deep learning-based multi-environment whole-genome prediction system, implemented based on the aforementioned deep learning-based multi-environment whole-genome prediction method, comprising a feature extraction module, a feature fusion module, a convolutional neural network module, and a prediction regression module; the feature extraction module is used to extract features from genotype data and environmental data; the feature fusion module is used to fuse genotype and environmental features; the convolutional neural network module is used to extract features and learn representations from the fused features; and the prediction regression module is used to regress the features processed by the convolutional neural network module to predict the phenotypic value of the trait.

[0102] An electronic device is also provided, comprising: a memory, a processor, and a computer program, the computer program being stored in the memory and configured to be executed by the processor to implement the aforementioned deep learning-based multi-environment whole-genome prediction method.

[0103] Additionally, a computer-readable storage medium is provided, in which a computer program is stored, which is executed by a processor to implement the aforementioned deep learning-based multi-environment whole-genome prediction method.

[0104] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0105] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A multi-environment whole-genome prediction method based on deep learning, characterized in that: Includes the following steps: Step S1: Obtain a dataset including genotype, phenotype, and environmental type; Step S2: Construct a multi-environment whole-genome prediction model including a genotype feature extractor, an environment feature extractor, a feature fusion module, a convolutional neural network module, and a prediction regression module; The genotype feature extractor is used to extract features from genotype data; The environmental feature extractor extracts environmental information from the entire growth cycle by adding a learnable element to the Transformer encoder and interacting with other environmental elements of variable sequence length. The feature fusion module is used to fuse genotype and environmental features; The convolutional neural network module is used to extract features and learn representations from the fused features; The prediction regression module is used to regress the features passed through the convolutional neural network module in order to predict the phenotypic value of the trait. Step S3: Split the dataset into training and validation sets under various segmentation scenarios, and input them into the multi-environment whole genome prediction model to obtain the phenotypic value prediction results of crop traits; Step S4: Calculate the mean squared error between the predicted result and the phenotypic values ​​in the validation set as the phenotypic prediction loss, and update the parameters of the prediction model by optimizing this loss.

2. The multi-environment whole-genome prediction method based on deep learning according to claim 1, characterized in that: Step S2 describes a genotype feature extractor that maps each locus of the gene data to a high-dimensional space to extract features from the genotype data. ； in, For the process Genotype data features after feature transformation; This is genotype data.

3. The multi-environment whole-genome prediction method based on deep learning according to claim 1, characterized in that: The environmental feature extractor described in step S2 uses a Transformer encoder, which consists of a multi-head self-attention mechanism and an alternating MLP block layer. By treating environmental data at different times as elements and prepending a learnable element in the environmental data, the environmental information of the entire growth cycle is aggregated through the self-attention mechanism in each layer of the Transformer.

4. The multi-environment whole-genome prediction method based on deep learning according to claim 3, characterized in that: In each layer of the Transformer encoder, the expression for interacting learnable elements with other elements is: ； ； ； ； In the formula, Input features; E represents environmental data; For learnable embeddings; For positional embedding; LN is for normalization; For multi-head attention operations; As an intermediate feature; The encoder output; learnable embeddings At the output of the converter encoder, corresponding to MLP is a multilayer perceptron module containing two layers. This is a characteristic of environmental data.

5. The multi-environment whole-genome prediction method based on deep learning according to claim 4, characterized in that: By calculating the self-attention of each element, we can extend it to obtain the self-attention mechanism. The expression for calculating the self-attention of each element is as follows: ； ； In the formula, H represents the input dimension, and Q, K, and V represent the query, key, and value, respectively. This is the activation function.

6. The multi-environment whole-genome prediction method based on deep learning according to claim 1, characterized in that: The various segmentation scenarios mentioned in step S3 include: a normal scenario, a genotype-specific scenario, and an environment-specific scenario; the normal scenario uses five-fold cross-validation to segment all combinations; the genotype-specific scenario uses five-fold segmentation based on the unique identifier of the genotype; and the environment-specific scenario uses five-fold segmentation based on the unique identifier of the environment.

7. The multi-environment whole-genome prediction method based on deep learning according to claim 1, characterized in that: Step S4 also employs CNN feature map visualization technology to perform interpretability analysis on the model, providing practical guidance for intelligent planting management.

8. A deep learning-based multi-environment whole-genome prediction system, applicable to the implementation of the deep learning-based multi-environment whole-genome prediction method as described in any one of claims 1 to 7, characterized in that: The system includes a feature extraction module, a feature fusion module, a convolutional neural network module, and a prediction and regression module; The feature extraction module is used to extract features from genotype data and environmental data; The feature fusion module is used to fuse genotype and environmental features; The convolutional neural network module is used for feature extraction and representation learning of the fused features; The prediction regression module is used to regress the features passed through the convolutional neural network module in order to predict the phenotypic value of the trait.

9. An electronic device, comprising: A memory, a processor, and a computer program, the computer program being stored in the memory and configured to be executed by the processor to implement a deep learning-based multi-environment whole-genome prediction method as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program that is executed by a processor to implement a deep learning-based multi-environment whole-genome prediction method as described in any one of claims 1-7.

Images