A genome prediction method and system based on double attention gate fusion DAGF

Abstract

The application provides a kind of based on double attention gate fusion DAGF genome prediction method and system, by adopting pre-training-fusion three-section architecture: first, in each source environment, independent pre-training is carried out using multi-scale one-dimensional convolution and hybrid pooling attention module, and local and global features of genome are extracted;Second, fix the feature extraction module on a small amount of data in the target environment, fine-tune the last two fully connected layers, realize fast environment adaptation;Finally, through the channel and spatial double attention mechanism, the features of each source environment are weighted, and the multi-head representation is learned by using the learnable gate fusion, so as to adaptively integrate multi-environment knowledge and suppress negative transfer. The adaptive feature extraction of multi-source environment genome data, environment-specific adaptation and intelligent fusion function are realized. It is suitable for predicting the cross-environment genome of complex quantitative traits with low genetic force and strong environmental interaction in crop breeding, improving the accuracy, robustness and computational efficiency of cross-environment prediction.

Description

Technical Field

[0001] This invention belongs to the field of computational genomics and bioinformatics, specifically relating to a genome prediction method and system based on dual attention gating fusion DAGF. Background Technology

[0002] Genomic prediction (GP) is one of the core technologies in modern crop breeding. Based on genome-wide molecular markers (such as single nucleotide polymorphisms, SNPs), this method uses statistical or machine learning models to achieve early and efficient prediction of unknown phenotypes (especially complex quantitative traits) in individual plants. GP technology significantly shortens the breeding cycle and reduces costs, making it a crucial driving force for the development of precision breeding.

[0003] In traditional genome prediction frameworks, models are typically trained and validated on single, homogeneous environmental data. However, the actual phenotype of a crop is determined by genotype (G), environment (E), and the interaction between genotype and environment (G×E). In practical breeding, two major challenges are frequently encountered: 1) scarce or completely missing observational data for the target promotion environment (e.g., new regions, new climatic conditions); and 2) the same genotype may exhibit significantly different phenotypes under different environmental conditions, i.e., a significant G×E effect exists. Therefore, how to utilize genotype-phenotype data accumulated in multiple historical environments (source environments) to construct a robust and efficient model to accurately predict the performance of individuals in entirely new target environments (i.e., "cross-environment genome prediction" or "genome cross-environment transfer learning") has become a research challenge and frontier in the field of crop breeding.

[0004] Existing technologies have the following limitations when dealing with cross-environment genome prediction problems: (1) Insufficient ability to model environmental heterogeneity: Although traditional multi-environment models (such as multi-environment GBLUP) can analyze multi-environment data at the same time, their modeling of environmental effects is usually relatively simple (considered as fixed or random effects), making it difficult to fully capture complex, nonlinear environment-specific patterns and high-order interaction relationships between genotypes and the environment; (2) Coarse-grained and inefficient transfer learning strategies: Some studies have attempted to apply transfer learning to cross-environment prediction, for example, by pre-training the model in the source environment and then fine-tuning it on a small sample of the target environment. However, such methods usually lack explicit, fine-grained differentiation and fusion mechanisms for shared and specific information between environments. They often transfer and adjust model parameters as a whole, ignoring the dynamic changes in the importance of genomic features in different environments, resulting in an inaccurate transfer process and susceptibility to negative transfer (i.e., source environment knowledge interfering with target environment prediction); (3) Insufficient utilization of the features of high-dimensional genomic data: Genomic data has the characteristics of high dimension and sparseness. Existing models often directly splice or simply weight environmental features when fusing multi-environment data, failing to establish an effective mechanism for adaptively selecting the most discriminative key features for the current prediction task (especially specific environments), thus affecting the model's prediction accuracy and generalization ability; (4) Poor computational efficiency and scalability: Although some complex deep learning models have strong representation capabilities, they have a large number of parameters and require massive amounts of data for training. In actual breeding scenarios, the sample size of a single environment is often limited, which can easily lead to model overfitting. At the same time, processing multi-environment, high-dimensional genomic data requires high computational resources, limiting its application in large-scale breeding populations.

[0005] In summary, existing genome prediction methods still have significant shortcomings when dealing with complex prediction tasks across multiple and cross-environment environments, particularly in terms of the precision of environment-specific modeling, the intelligence of cross-environment knowledge transfer, the adaptability of high-dimensional feature selection, and overall computational efficiency. Therefore, there is an urgent need in this field for new models and methods that can adaptively integrate multi-environment information, accurately distinguish between shared and unique features, and efficiently perform cross-environment predictions. This would improve the accuracy and robustness of predicting the performance of complex quantitative traits in crop breeding under different environments, and provide a more powerful computational tool for cross-regional and cross-climate breeding decisions. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a genome prediction method and system based on dual attention gating fusion (DAGF) for adaptive feature extraction, environment-specific adaptation and intelligent fusion of multi-source environmental genome data.

[0007] The technical solution adopted by this invention to solve the above-mentioned technical problems is as follows: a genome prediction method based on dual attention-gated fusion DAGF, comprising the following steps: S1: Pre-train the genotype-phenotype data for each source environment, and use multi-layer one-dimensional convolutional neural networks and hybrid pooling attention to extract local and global features of SNP sequences to obtain the pre-trained model; S2: When there is a small amount of labeled data in the target environment, perform lightweight adaptation on the pre-trained model to obtain a fine-tuned model; S3: Integrates fine-tuned models from different source environments, and achieves intelligent fusion of multi-source knowledge through dual attention mechanism and gating fusion; S4: By integrating and fusing the features through a fully connected layer, the phenotypic prediction of the target environment is output.

[0008] According to the above scheme, the specific steps in step S1 are as follows: S11: Stacked convolutions using kernels of different sizes are used to capture local genetic patterns at different scales in SNP sequences and increase the number of feature channels layer by layer to enhance expressive power. S12: Introducing a group attention mechanism, the feature channels are divided into several groups to reduce computational complexity and capture the correlation of features within the group; average pooling and max pooling are combined within each group; channel weights are generated through 1×1 convolution, and each group of features is weighted to highlight key SNP sites, and element-wise gating is used to further refine the attention weights and improve the accuracy of feature selection.

[0009] Furthermore, step S1 also includes a flattening operation, and outputs the trained feature representation through a fully connected layer and a Dropout layer.

[0010] According to the above scheme, the specific steps in step S2 are as follows: S21: Fix the parameters of all convolutional and attention layers in the pre-training phase, and preserve the learned general genomic feature representations; S22: Replace and retrain the top fully connected layer to adapt to the phenotypic distribution characteristics of the target environment.

[0011] Furthermore, in step S2, while preserving cross-environment shared features, the model quickly adapts to environment specificity by reducing the number of updated parameters, balancing the model's fitting ability and generalization performance, and avoiding overfitting.

[0012] According to the above scheme, the specific steps in step S3 are as follows: S31: Extract and copy the features generated by the convolutional layer, HPA projection layer, attention layer, and the first fully connected layer from each fine-tuned model; S32: CBAM1d dual attention weighting; S33: Head-to-head gating fusion.

[0013] According to the above scheme, the specific steps in step S32 are as follows: First, channel attention is calculated, and the feature channels of different models are weighted to highlight the feature channels that are important for predicting the target environment. Then, spatial attention is calculated, and weights are calculated on the spatial dimension of the features to enhance the response of key genetic loci.

[0014] According to the above scheme, the specific steps in step S33 are as follows: Each feature representation corresponding to a source environment is regarded as a head. The contribution ratio of each head in the final prediction is adjusted by a learnable dynamic gating weight, thereby achieving adaptive fusion of cross-environment knowledge.

[0015] A genome prediction system based on dual attention-gated fusion DAGF The pre-training submodule is used to pre-train the genotype-phenotype data of each source environment. It uses a multi-layer one-dimensional convolutional neural network to extract the local and global features of the SNP sequence to obtain the pre-trained model. The fine-tuning submodule is used to perform lightweight adaptation on the pre-trained model when there is a small amount of labeled data in the target environment, so as to obtain a fine-tuned model; The fusion submodule is used to integrate fine-tuned models from different source environments, and achieves intelligent fusion of multi-source knowledge through dual attention mechanism and gating fusion; The output submodule is used to integrate and fuse the features through a fully connected layer and output the phenotypic prediction value of the target environment.

[0016] A computer memory storing a computer program executable by a computer processor, the computer program performing a genome prediction method based on dual attention-gated fusion DAGF.

[0017] The beneficial effects of this invention are as follows: 1. This invention discloses a genome prediction method and system based on dual attention-gated fusion (DAGF). The method employs a three-stage architecture of "pre-training-fine-tuning-fusion": First, multi-scale one-dimensional convolutional and hybrid pooling attention modules are independently pre-trained in each source environment to extract local and global genomic features. Second, the underlying parameters are fixed and the top-level network is fine-tuned on a small amount of data from the target environment to achieve rapid environment adaptation. Finally, the features of each source environment are weighted through a dual attention mechanism of channel and space, and a learnable gated fusion multi-head representation is used to adaptively integrate multi-environment knowledge and suppress negative transfer. This achieves adaptive feature extraction, environment-specific adaptation, and intelligent fusion of multi-source environmental genomic data.

[0018] 2. This invention addresses the problems of insufficient environmental heterogeneity modeling, extensive transfer learning strategies, low efficiency in utilizing high-dimensional features, and poor computational scalability in existing methods for cross-environment genome prediction of crops. It significantly improves the accuracy and robustness of cross-environment prediction, and is particularly suitable for complex quantitative traits with low heritability and strong environmental interactions, providing an efficient computational tool for cross-regional crop breeding.

[0019] 3. This invention is applicable to cross-environmental genome prediction of complex quantitative traits in crop breeding, and can effectively improve the accuracy, robustness and computational efficiency of prediction.

[0020] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of an embodiment of the present invention.

[0023] Figure 2 This is the overall architecture and flowchart of each module of the DAGF algorithm in this embodiment of the invention.

[0024] Figure 3 This is a comparison chart of the PCC of the pre-trained infrastructure of this invention with eight other models in the BJ single environment of the corn NCII data.

[0025] Figure 4 This is a comparison chart of PCC results for cross-environment prediction of DTT traits in an embodiment of the present invention.

[0026] Figure 5 This is a comparison chart of PCC results for cross-environment prediction of pH properties in an embodiment of the present invention.

[0027] Figure 6 This is a comparison chart of PCC results for cross-environment prediction of EW traits in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0029] Example 1 See Figure 1 The specific steps of a genome prediction method based on dual attention-gated fusion DAGF are as follows: 1. Pre-training phase For each source environment's genotype-phenotype data, independent pre-trained modules were constructed. Each module employed a multi-layer one-dimensional convolutional neural network (1D-CNN) to extract local and global features of the SNP sequences. Specifically, this included: 1) Multi-scale one-dimensional convolutional layer: Stacked convolutions with three different sizes of kernels (9, 5, and 3) are used to capture local genetic patterns at different scales in SNP sequences, and the number of feature channels is increased layer by layer to enhance expressive power. 2) Hybrid Pooling Attention Module (HPA1d): This module introduces a grouped attention mechanism on top of convolutional features, dividing the feature channels into several groups (e.g., 4 groups) to reduce computational complexity and capture the correlation between features within each group. Within each group, average pooling (capturing global trends) and max pooling (capturing local significant peaks) are combined to avoid information loss caused by single pooling. Subsequently, channel weights are generated through 1×1 convolutions to weight each group of features to highlight key SNP sites, and element-wise gating is used to further refine the attention weights, improving feature selection accuracy. This module is followed by a flattening operation and outputs the training feature representation through fully connected layers (FC layers) and dropout layers.

[0030] 2. Fine-tuning stage When the target environment has a small amount of labeled data, perform lightweight adaptation on the pre-trained model. Specific steps include: 1) Fix the parameters of all convolutional and attention layers in the pre-training phase, and retain the learned general genomic feature representations; 2) Replace and retrain only the top fully connected layer (including Dropout) to adapt it to the phenotypic distribution characteristics of the target environment; 3) This design retains the shared features across environments, and quickly adapts to environment specificity with very few parameter updates, effectively balancing the model's fitting ability and generalization performance, and avoiding overfitting.

[0031] 3. Fusion Phase Integrating the outputs of fine-tuned models from different source environments, intelligent fusion of multi-source knowledge is achieved through a dual attention mechanism and gating fusion, specifically including: 1) Feature duplication: Extract and duplicate the features generated by the convolutional layer, HPA projection layer, attention layer, and the first fully connected layer from each fine-tuned model; 2) CBAM1d Dual Attention Weighting: First, channel attention is calculated to weight the feature channels of different models (corresponding to different SNP features) to highlight the feature channels that are important for predicting the target environment; then, spatial attention is calculated to calculate the weights on the spatial dimension of the features (corresponding to SNP sites) to strengthen the response of key genetic sites. 3) Head-by-head gating fusion: The feature representation corresponding to each source environment is regarded as a "head". The contribution ratio of each head in the final prediction is adjusted by a learnable dynamic gating weight, so as to achieve adaptive fusion of cross-environment knowledge, effectively suppress negative transfer, and improve the prediction specificity of the model for the target environment.

[0032] Finally, the fused features are integrated by a fully connected layer to output the phenotypic prediction of the target environment.

[0033] This embodiment employs a three-stage architecture of "pre-training-fine-tuning-fusion": First, multi-scale one-dimensional convolution and hybrid pooling attention modules are independently pre-trained in each source environment to extract local and global genomic features. Second, the underlying parameters are fixed and the top-level network is fine-tuned on a small amount of data from the target environment to achieve rapid environment adaptation. Finally, the features of each source environment are weighted through a dual channel and spatial attention mechanism, and learnable gating is used to fuse multi-head representations, thereby adaptively integrating multi-environment knowledge and suppressing negative transfer. This achieves adaptive feature extraction, environment-specific adaptation, and intelligent fusion of multi-source environmental genomic data.

[0034] Example 2 The steps in this embodiment are the same as in Embodiment 1, the difference being that each step is applied to a specific instance. See also Figure 2 The figure shows: (a) overall model architecture; (b) pre-trained model framework; (c) fine-tuned model framework; (d) fusion model framework; specifically including the following steps: 1. Experimental Setup Development environment: PyTorch 2.1.2; Python 3.10 (Ubuntu 22.04); CUDA 11.8; GPU: RTX 3090.

[0035] 2. Test data and preprocessing: (1) Pre-trained basic framework single-environment test data: 5506 samples of corn NCIIZ in BJ environment, 10000 SNP labels; (2) DAGF migration architecture cross-environment test data: Based on 10,000 SNP markers of the maize NCII population (sample size n=5506) and their phenotypic values ​​in five environments, 20 random partitions were performed; each partition yielded five groups of corresponding genotype and phenotypic data, and a fine-tuning dataset was generated at the same time, whose genotype data matched the phenotypic values ​​of the five environments, and whose sample size was approximately 30% of the sample size of each environment; (3) Data preprocessing: genotype matrix columns are standardized (μ=0, σ²=1), and phenotypic values ​​are normalized to (-1, 1).

[0036] 3. Parameter Settings (1) Pre-training infrastructure parameters: The basic number of channels in the convolutional layer is 1, the kernel size is 9, 5, and 3 respectively, the stride is 1, the number of neurons in the fully connected layer is 64, the dropout probability is 0.3, the batch size is 64, the optimizer is Adam, and the initial learning rate is 1e-5; the HPA1d module is set to 8 channels and 4 groups, and the element-level gating function is enabled. A gating value with the same shape as the feature is generated by 1×1 convolution.

[0037] (2) Cross-environment DAGF migration architecture parameters: Pre-training phase: learning rate 1e-3, optimizer Adam, number of neuron units 64, batch size 64, dropout probability 0.3, HPA1d module as before; Fine-tuning phase: learning rate 5e-5, optimizer Adam, number of neuron units 64, batch size 64, dropout probability 0.3; Fusion phase: learning rate 1e-3, optimizer Adam, number of neuron units 64, batch size 64, dropout probability 0.3.

[0038] 4. Verification Plan (1) Prediction within a single environment: For each environment's valid samples, 20 independent random partitions are performed, using a two-layer dataset partitioning strategy: first, the samples are divided into training and test sets at a ratio of 8:2, and then the actual training and validation sets are split from the training set at a ratio of 9:1. Random seeds are dynamically generated to ensure the randomness and repeatability of the partitions. During the model training phase, the Adam optimizer and mean squared error loss function are used, combined with a dynamic learning rate scheduling strategy and early stopping. The Pearson correlation coefficient (PCC) is used to evaluate the model performance. (2) Cross-environment migration prediction: Leave-one-out method is used, with five environments rotating as target environments and the remaining four as source domains. The entire process is divided into three stages, with five-fold cross-validation used in each stage. The model with the smallest mean squared error (MSE) is selected as the final model. The prediction PCC of the target environment is used as the main evaluation index.

[0039] 5. Results Analysis (1) Prediction within a single environment: For the EW trait prediction task on the maize NCII dataset in a BJ single-environment environment, the prediction performance of nine different modeling methods was compared. See [link to relevant documentation]. Figure 3 This is a comparison of the PCC (Predictive Comparison) of the pre-trained basic framework on the single-environment (BJ) corn NCII data of the example with eight other models, including rrblup, CNN, DNNGP, convolution + Mamba2 + VAE fusion, three-layer convolution, three-layer convolution + local attention + fully connected, VAE convolution parallel framework, and sparse coding + Mamba2 + residual processing + fully connected. The average Pearson correlation coefficient (PCC) was used as the evaluation metric (a PCC closer to 1 indicates better prediction performance). Experimental results show that the pre-trained basic framework model achieved the highest average PCC value of 0.7494, significantly outperforming all other compared models, indicating that this model possesses the strongest feature extraction and association mapping capabilities in single-environment phenotypic prediction tasks.

[0040] (2) Cross-environment prediction: DTT characteristics: See Figure 4 The horizontal axis represents the target environment, and the predictive performance of each model shows a clear gradient across different environments. In the JL (jl) and LN (ln) environments, the PCC values ​​of all models are generally higher than 0.8, followed by the HN (hn) environment, and the lowest in the HB (heb) environment, reflecting the environmental heterogeneity of the heritability of this trait. The DAFG and finetune_data_pre_dict architectures, i.e., models directly trained on a pre-trained framework based on 30% of the target environment fine-tuned data, perform best in most environments. For example, in the jl environment, their PCC values ​​reach 0.892 and 0.883, respectively. Compared to benchmark methods such as rrblup and pre_dict (models trained on a pre-trained framework based on full-source domain data and target domain fine-tuned data), the model fine-tuned with 30% of the target environment data significantly improves the stability and accuracy of cross-environment predictions, validating the gain effect of the fine-tuning strategy on predictions in low heritability environments.

[0041] pH properties: See Figure 5The horizontal axis represents the target environment. Overall predictive performance across environments shows that the PCC values ​​of all models are generally higher than 0.8 in the BJ (bj) and HN (hn) environments, while they slightly decrease in the JL (jl) and LN (ln) environments. This trend indicates that the overall degree of environmental regulation of plant height phenotypic expression is relatively low, but significant environment specificity still exists. At the model level, DAFG shows the most stable performance and a significant advantage among all models, demonstrating that transfer learning strategies can effectively uncover environment-specific genetic signals.

[0042] EW characteristics: see Figure 6 The horizontal axis represents the target environment. Predictive performance varies significantly across environments. The BJ (bj) and HB (heb) environments show better performance (PCC≈0.6), while the JL (jl) and LN (ln) environments show a significant decline (PCC<0.5), reflecting the high impact of environmental noise on yield traits and the difficulty of prediction. DAFG performs best among all models, validating the significant value of the "pre-training-fine-tuning" transfer strategy in cross-environment prediction of low heritability traits.

[0043] This embodiment addresses the problems of insufficient environmental heterogeneity modeling, extensive transfer learning strategies, low efficiency in utilizing high-dimensional features, and poor computational scalability in existing methods for cross-environment genome prediction of crops. It significantly improves the accuracy and robustness of cross-environment prediction, and is particularly suitable for complex quantitative traits with low heritability and strong environmental interactions, providing an efficient computational tool for cross-regional crop breeding.

[0044] It should be understood that the sequence number of each step in the above embodiments does not imply 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.

[0045] Example 3 This embodiment is used to implement the principle of the above method embodiment to construct a genome prediction system based on dual attention gating fusion DAGF, including a pre-training submodule, a fine-tuning submodule, a fusion submodule and an output submodule.

[0046] The pre-training submodule is used to pre-train the genotype-phenotype data of each source environment. It uses a multi-layer one-dimensional convolutional neural network to extract the local and global features of the SNP sequence to obtain the pre-trained model. The fine-tuning submodule is used to perform lightweight adaptation on the pre-trained model when there is a small amount of labeled data in the target environment, so as to obtain a fine-tuned model; The fusion submodule is used to integrate fine-tuned models from different source environments, and achieves intelligent fusion of multi-source knowledge through dual attention mechanism and gating fusion; The output submodule is used to integrate and fuse the features through a fully connected layer and output the phenotypic prediction value of the target environment.

[0047] Each submodule is mainly used to implement the various steps of the method implementation, which will not be elaborated here.

[0048] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0049] This embodiment also includes a processor, a communication interface, a memory, and a communication bus; wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of a genome prediction method based on dual attention gating fusion DAGF.

[0050] This embodiment also provides a computer-readable storage medium storing executable instructions that, when executed by a processor, enable the processor to implement a genome prediction method based on dual attention-gated fusion (DAGF).

[0051] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0052] Furthermore, this application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0053] This application is described with reference to the flowchart of the method and computer program product according to Embodiment 1 and the block diagram of the device (system) according to Embodiment 3. It should be understood that each step or block in the flowchart or block diagram, as well as combinations of steps or blocks in the flowchart or block diagram, can be implemented by computer program instructions.

[0054] These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which are executable by the processor of the computer or other programmable data processing device, produce instructions for implementing the process. Figure 1 One or more processes or boxes Figure 1 A genome prediction system based on dual attention-gated fusion (DAGF) that specifies functions in one or more boxes.

[0055] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes or boxes Figure 1 The function specified in one or more boxes.

[0056] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes or boxes Figure 1 The steps of a genome prediction method based on dual attention-gated fusion (DAGF) are specified in one or more boxes.

[0057] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A genome prediction method based on dual attention-gated fusion DAGF, characterized in that: Includes the following steps: S1: Pre-train the genotype-phenotype data for each source environment, and use multi-layer one-dimensional convolutional neural networks and hybrid pooling attention to extract local and global features of SNP sequences to obtain the pre-trained model; S2: When there is a small amount of labeled data in the target environment, perform lightweight adaptation on the pre-trained model to obtain a fine-tuned model; S3: Integrates fine-tuned models from different source environments, and achieves intelligent fusion of multi-source knowledge through dual attention mechanism and gating fusion; S4: By integrating and fusing the features through a fully connected layer, the phenotypic prediction of the target environment is output.

2. The genome prediction method based on dual attention-gated fusion DAGF according to claim 1, characterized in that: The specific steps in step S1 are as follows: S11: Stacked convolutions using kernels of different sizes are used to capture local genetic patterns at different scales in SNP sequences and increase the number of feature channels layer by layer to enhance expressive power. S12: Introduce a group attention mechanism to divide the feature channels into several groups to reduce computational complexity and capture the correlation of features within the group; Within each group, average pooling and max pooling are combined; channel weights are generated through 1×1 convolutions, and the features in each group are weighted to highlight key SNP sites. Element-wise gating is then used to further refine the attention weights. Improve the accuracy of feature selection.

3. The genome prediction method based on dual attention-gated fusion DAGF according to claim 2, characterized in that: Step S1 further includes a flattening operation, and outputs the trained feature representation through a fully connected layer and a Dropout layer.

4. The genome prediction method based on dual attention-gated fusion DAGF according to claim 1, characterized in that: The specific steps in step S2 are as follows: S21: Fix the parameters of all convolutional and attention layers in the pre-training phase, and preserve the learned general genomic feature representations; S22: Replace and retrain the top fully connected layer to adapt to the phenotypic distribution characteristics of the target environment.

5. A genome prediction method based on dual attention-gated fusion DAGF according to claim 4, characterized in that: In step S2, while preserving cross-environment shared features, the model quickly adapts to environment specificity by reducing the number of updated parameters, balancing the model's fitting ability and generalization performance, and avoiding overfitting.

6. The genome prediction method based on dual attention-gated fusion DAGF according to claim 1, characterized in that: The specific steps in step S3 are as follows: S31: Extract and copy the features generated by the convolutional layer, HPA projection layer, attention layer, and the first fully connected layer from each fine-tuned model; S32: CBAM1d dual attention weighting; S33: Head-to-head gating fusion.

7. The genome prediction method based on dual attention-gated fusion DAGF according to claim 1, characterized in that: The specific steps in step S32 are as follows: First, channel attention is calculated, and the feature channels of different models are weighted to highlight the feature channels that are important for predicting the target environment. Then, spatial attention is calculated, and weights are calculated on the spatial dimension of the features to enhance the response of key genetic loci.

8. The genome prediction method based on dual attention-gated fusion DAGF according to claim 1, characterized in that: The specific steps in step S33 are as follows: Each feature representation corresponding to a source environment is regarded as a head. The contribution ratio of each head in the final prediction is adjusted by a learnable dynamic gating weight, thereby achieving adaptive fusion of cross-environment knowledge.

9. A genome prediction system based on dual attention-gated fusion DAGF, characterized in that: The pre-training submodule is used to pre-train the genotype-phenotype data of each source environment. It uses a multi-layer one-dimensional convolutional neural network to extract the local and global features of the SNP sequence to obtain the pre-trained model. The fine-tuning submodule is used to perform lightweight adaptation on the pre-trained model when there is a small amount of labeled data in the target environment, so as to obtain a fine-tuned model; The fusion submodule is used to integrate fine-tuned models from different source environments, and achieves intelligent fusion of multi-source knowledge through dual attention mechanism and gating fusion; The output submodule is used to integrate and fuse the features through a fully connected layer and output the phenotypic prediction value of the target environment.

10. A computer memory, characterized in that: It contains a computer program that can be executed by a computer processor, which performs a genome prediction method based on dual attention-gated fusion DAGF as described in any one of claims 1 to 8.

Images