Tomato breeding prediction and design method based on bidirectional generative model

Abstract

The invention discloses a tomato breeding prediction and design method based on a bidirectional generation model. The method comprises the following steps: firstly, constructing a tomato genotype and phenotype comprehensive data set, and carrying out feature correlation analysis by utilizing an attention mechanism; and then constructing a bidirectional deep learning model containing a forward prediction and reverse design path. The forward network accurately predicts phenotypes from genotypes by adopting a self-attention mechanism, the reverse network reversely deduces virtual genotypes from target phenotypes through transposition convolution, and cyclic consistency and linkage imbalance constraints are introduced for joint training. According to the method, the problem of nonlinear mapping between genotypes and phenotypes is effectively solved, two-way intercommunication between character prediction and parent design is realized, and the precision breeding efficiency of tomatoes is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics and artificial intelligence, and in particular to a method for predicting and designing crop genotypes and agronomic traits based on deep learning and bidirectional generative models. Background Technology

[0002] With the development of high-throughput sequencing technology, crop breeding has entered the era of intelligent breeding driven by big data. Genomic selection (GS) technology, by constructing predictive models and predicting phenotypes based on genotype data, has greatly accelerated the breeding process. However, analysis of existing technologies reveals that current breeding models still have the following limitations:

[0003] First, most traditional models are based on linear assumptions, making it difficult to capture complex nonlinear epistasis between genes, thus limiting their prediction accuracy for complex traits controlled by multiple genes. Second, existing deep learning models mostly focus on unidirectional prediction, making it difficult to achieve "reverse design," that is, to deduce the desired genotype combination from the ideal breeding goal.

[0004] Despite the excellent performance of Generative Adversarial Networks (GANs) in image generation, there are significant technical obstacles to their direct application in genomic data generation: the generated genotype data often ignores the biologically specific linkage disequilibrium (LD) pattern, resulting in biologically unreasonable gene sequences that exhibit "pseudo-correlation" and cannot be used to guide actual hybridization breeding programs.

[0005] Therefore, there is an urgent need for an intelligent breeding method that can accurately capture epistatic effects, ensure the biological rationality of the generated data, and support bidirectional mapping between genotype and phenotype. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting and designing tomato genotypes and phenotypes based on a bidirectional generative model, which aims to solve the problems of existing technologies being unable to simultaneously capture epistatic effects and biological constraints, as well as lacking bidirectional design capabilities.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for predicting and designing tomato genotypes and phenotypes based on a bidirectional generative model includes the following steps:

[0009] Step S1: Obtain the genotype data and corresponding agronomic phenotypic data of the tomato population, perform one-hot encoding on the genotype data, and normalize the agronomic phenotypic data.

[0010] Step S2: Construct a breeding deep learning model, which includes a forward prediction network, a backward design network, and a shared latent feature mapping module;

[0011] Step S3: Train the breeding deep learning model using the processed genotype data and agronomic phenotypic data. During the training process, the cycle consistency loss function, adversarial loss function, and linkage disequilibrium constraint loss function are introduced.

[0012] Step S4: Execute model inference application; specifically including:

[0013] In response to the prediction command, the genotype data of the tomato material to be tested is input into the trained positive prediction network, and the predicted agronomic phenotypic data is output.

[0014] In response to design instructions, the target agronomic trait phenotypic data are input into the trained reverse design network, which outputs the target genotype data of the design.

[0015] Furthermore, the positive prediction network includes a genotype encoder and a phenotype decoder; the genotype encoder adopts a multi-layer self-attention mechanism structure to capture epistatic effects between gene loci and extract genotype feature vectors; the phenotype decoder adopts a fully connected layer structure to map genotype feature vectors to phenotype prediction values.

[0016] Furthermore, the reverse design network includes a phenotypic encoder and a genotype generator; the phenotypic encoder is used to extract phenotypic feature vectors; the genotype generator adopts a transposed convolutional structure to reconstruct the phenotypic feature vectors into a genotype probability distribution matrix.

[0017] Furthermore, the formula for calculating the chain imbalance constraint loss function is as follows: .

[0018] in, This represents the generated genotype matrix. This represents the linkage disequilibrium correlation coefficient matrix pre-calculated based on real population data. This represents the Frobenius norm.

[0019] Furthermore, the agronomic phenotypic data include at least the soluble solids content of the fruit, the lycopene content, and the yield per plant.

[0020] The present invention also provides a tomato breeding prediction and design device based on a bidirectional generative model, including a data acquisition module, a preprocessing module, a model building module, a model training module, and a prediction and design module.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] Bidirectional mapping and functional expansion: This invention innovatively introduces a cyclic consistency architecture, which breaks through the limitation of unidirectional prediction in traditional models, and realizes accurate prediction from genotype to phenotype and reverse design from phenotype to genotype, providing an integrated "prediction + design" solution for molecular breeding.

[0023] Effective capture of epistatic effects: By utilizing the multilayer self-attention mechanism in the forward network, it is possible to effectively model long-distance interactions and nonlinear epistatic effects between gene loci, which significantly improves the prediction accuracy of complex agronomic traits.

[0024] Biological rationality constraint mechanism: In response to the problem that generative models are prone to generating invalid data, this invention introduces the LD constraint loss function, which forces the genotypes generated by the model to follow the genetic linkage law of natural populations, ensuring that the genotypes designed in reverse are biologically rational and operable, and have extremely high practical value. Attached Figure Description

[0025] Figure 1 This is a flowchart of the tomato breeding prediction and design method based on a bidirectional generative model provided in Embodiment 1 of the present invention;

[0026] Figure 2 This is a heatmap of gene locus attention weights captured by the model in this embodiment of the invention, demonstrating the interpretability of the model;

[0027] Figure 3 This is a schematic diagram of the overall architecture of the deep learning model constructed in Embodiment 2 of the present invention, showing the connection relationship between the forward prediction network, the reverse design network, and the shared latent feature mapping module;

[0028] Figure 4 This is a scatter plot of the prediction results of the model on the test set in Embodiment 3 of the present invention, which shows the degree of fit between the predicted phenotypic values ​​and the true phenotypic values. Detailed Implementation

[0029] Example 1

[0030] This embodiment provides a tomato breeding prediction and design method based on a bidirectional generative model. For example... Figure 1 As shown, the method first involves data acquisition and construction, namely, collecting tomato germplasm resource data from natural populations and constructing a comprehensive dataset containing genotype data and corresponding phenotypic data.

[0031] Subsequently, the raw data underwent preprocessing and feature analysis, including data cleaning, missing value imputation, and numerical normalization. During this process, to improve the model's training efficiency and interpretability, this embodiment also conducted a preliminary analysis of the association between genotype loci and phenotypic traits. Figure 2 The feature correlation heatmap generated in this embodiment shows that the horizontal and vertical axes represent different gene loci or trait dimensions, respectively, and the shades of the square colors represent the correlation strength or the model's focus weight. Darker colored areas indicate a significant and strong association between these features. This visualized feature analysis allows verification of whether the model has focused on biologically significant key sites in subsequent steps, ensuring the scientific rigor of the method.

[0032] After data preparation, a deep learning architecture comprising a forward prediction network and a backward design network was constructed, and a cycle consistency mechanism was introduced. The model was iteratively trained using the collected dataset, employing a joint loss function including prediction loss, cycle consistency loss, and LD constraint loss to optimize model parameters. Finally, prediction and design were performed: in the forward prediction stage, the genotype data of the tomato material to be tested was input, and the predicted phenotypic values ​​were output through the forward network; in the backward design stage, the ideal breeding target phenotypic values ​​were input, and the corresponding virtual genotypes were generated through the backward network, thereby guiding the selection of breeding parents or the formulation of gene editing programs.

[0033] Example 2

[0034] This embodiment details the specific architecture of the deep learning model mentioned in Embodiment 1. For example... Figure 3 As shown, the breeding deep learning model constructed in this embodiment mainly includes a forward prediction network, a backward design network, and a shared latent feature mapping module.

[0035] First, the forward prediction network aims to achieve a mapping from genotype to phenotype. This network comprises a genotype encoder and a phenotype decoder. The genotype encoder employs a Transformer-based multi-head self-attention mechanism, which effectively captures long-distance dependencies and epistatic effects between different SNP sites. The phenotype decoder consists of fully connected layers, used to receive encoded features and output predicted phenotypic values.

[0036] Secondly, the reverse design network aims to reconstruct phenotypes from phenotypes. This network includes a phenotype encoder and a genotype generator, where the genotype generator specifically employs a transposed convolutional neural network (Deconv) structure. This structure can upsample low-dimensional latent features to restore them to a high-dimensional genotype probability distribution matrix, thereby generating the corresponding genotype sequence.

[0037] Finally, the model also includes a shared latent feature mapping module. This module aligns the latent space features of the forward and reverse pathways, ensuring semantic consistency between genotype and phenotypic features when mapped to the same dimension. This is a key component for achieving bidirectional cyclic prediction, guaranteeing accurate forward predictions while ensuring that the genotypes designed in reverse also accurately correspond to the target phenotype.

[0038] Example 3

[0039] This embodiment aims to verify the effectiveness of the above methods and models in actual tomato breeding data. The experiment selected 300 representative natural tomato populations, used resequencing technology to obtain whole-genome SNP variation information, and measured core agronomic traits such as single fruit weight and soluble solids content in the field. The dataset was divided into training, validation, and test sets in an 8:1:1 ratio, and the constructed bidirectional generative model was trained and optimized using the training set.

[0040] After the model is trained, it is validated on an independent test set. For example... Figure 4 As shown in the figure, this is a scatter plot of the model's predictive correlation on the test set. The horizontal axis represents the observed values ​​of the tomato trait, and the vertical axis represents the model's predicted values. The figure shows that the vast majority of data points are closely distributed around the diagonal, indicating a high degree of agreement between the predicted and actual values.

[0041] To further quantify and evaluate the performance of the model proposed in this invention, this embodiment also compares the proposed Bi-Genomic Model with traditional Linear Regression and Support Vector Machine (SVM) models. The evaluation metrics used are Pearson correlation coefficient (Pearson R) and root mean square error (RMSE), where a Pearson R value closer to 1 indicates a stronger correlation, and a smaller RMSE value indicates a lower error. The specific comparison results are shown in Table 1.

[0042] Table 1. Comparison of prediction performance of different models on tomato core traits.

[0043] Model Name Single fruit weight Pearson R Single fruit weight RMSE Soluble solids Pearson R soluble solids RMSE Linear Regression 0.65 0.15 0.58 0.18 SVM 0.78 0.11 0.72 0.14 Bi-Genomic 0.91 0.08 0.89 0.09

Claims

1. A tomato breeding prediction and design method based on a bidirectional generative model, characterized in that, Includes the following steps: Step S1: Obtain the genotype data and corresponding agronomic phenotypic data of the tomato population, perform one-hot encoding on the genotype data, and normalize the agronomic phenotypic data. Step S2: Construct a breeding deep learning model, which includes a forward prediction network, a backward design network, and a shared latent feature mapping module for aligning bidirectional features; Step S3: Use the processed genotype data and agronomic phenotypic data to train the breeding deep learning model end-to-end. The loss function in the training process includes the cycle consistency loss function, the adversarial loss function, and the linkage disequilibrium constraint loss function. Step S4: In response to the prediction instruction, the genotype data of the tomato material to be tested is input into the trained forward prediction network, and the predicted agronomic phenotypic data is output; and in response to the design instruction, the target agronomic phenotypic data is input into the trained reverse design network, and the target genotype data that conforms to the linkage disequilibrium constraint is output.

2. The method according to claim 1, characterized in that, The forward prediction network includes a genotype encoder and a phenotype decoder; the genotype encoder adopts a multi-layer self-attention mechanism structure and is configured to capture epistatic effects between gene loci and extract genotype feature vectors; the phenotype decoder adopts a fully connected layer structure and is configured to map genotype feature vectors to phenotype prediction values.

3. The method according to claim 1, characterized in that, The reverse design network includes a phenotypic encoder and a genotype generator; the phenotypic encoder is used to extract phenotypic feature vectors; the genotype generator adopts a transposed convolutional structure and is configured to reconstruct the phenotypic feature vectors into a genotype probability distribution matrix.

4. The method according to claim 1, characterized in that, The formula for calculating the chain imbalance constraint loss function is as follows: .in, This represents the generated genotype matrix. This represents the linkage disequilibrium correlation coefficient matrix pre-calculated based on real population data. This represents the Frobenius norm.

5. The method according to claim 1, characterized in that, The agronomic phenotypic data include at least the soluble solids content of the fruit, the lycopene content, and the yield per plant.

6. The method according to claim 1, characterized in that, The training process in step S3 specifically includes: The input genotype is mapped to the predicted phenotype using a forward prediction network, and then the predicted phenotype is reconstructed into the reconstructed genotype using a backward design network. The genotype cycle consistency loss is calculated. The input phenotype is mapped to the generated genotype using a reverse design network, and then the generated genotype is reconstructed into the reconstructed phenotype using a forward prediction network. The phenotype cycle consistency loss is calculated. By combining the adversarial loss function and the chain imbalance constraint loss function, the model parameters are jointly optimized using the gradient descent algorithm.

7. A tomato breeding prediction and design device based on a bidirectional generative model, characterized in that, include: The data acquisition module is used to acquire the genotype data and corresponding agronomic phenotypic data of the tomato population; The preprocessing module is used to perform one-hot encoding on the genotype data and normalize the agronomic phenotypic data. The model building module is used to build a breeding deep learning model that includes a forward prediction network, a backward design network, and a shared latent feature mapping module. The model training module is used to train the model using the processed data and calculate the cycle consistency loss, adversarial loss, and chain imbalance constraint loss. The prediction and design module is used to perform genotype-to-phenotype prediction and phenotype-to-genotype design.

8. The apparatus according to claim 7, characterized in that, The model training module is equipped with a linkage disequilibrium constraint unit, which stores a linkage disequilibrium correlation coefficient matrix pre-calculated based on real population data, and is used to apply biological constraints when generating genotypes in reverse.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 6.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 6.

Images