A multi-trait whole genome prediction method based on one-way cross trait attention
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- YANGZHOU UNIV
- Filing Date
- 2026-05-15
- Publication Date
- 2026-08-04
AI Technical Summary
[0005]本发明的目的是提供一种基于单向跨性状注意力的多性状全基因组预测方法，以解决以下问题：现有的深度学习MT模型主要采用传统的多任务学习范式，仅通过共享特征提取网络来隐式建模性状间的相关性，而各性状任务之间缺乏显式的信息交互机制
本发明摒弃了深度学习MT模型中仅依赖共享特征提取网络且无差别建模性状关联的通用多任务范式，通过构建单向跨性状注意力机制，实现了从辅助性状到目标性状的定向信息增强，有效规避了传统多任务学习中双向交互带来的冗余噪声干扰，从而显著提高了目标性状的GS预测精度。
Smart Images
Figure CN122511342A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of breeding technology, and in particular relates to a multi-trait genome-wide prediction method based on unidirectional cross-trait attention. Background Technology
[0002] Genomic selection (GS) utilizes genome-wide genetic markers for phenotypic prediction, effectively shortening breeding cycles and increasing genetic gains. Existing GS prediction models can be categorized into two types based on the number of traits analyzed: uni-trait (UT) models and multi-trait (MT) models. Compared to UT models, MT models are more effective at capturing complex genetic relationships between traits and typically exhibit superior predictive performance. To fully explore the genetic correlations between traits, current research is beginning to incorporate deep learning for MT prediction.
[0003] However, existing deep learning MT models (such as MtCro and DNAwhisper) generally adopt a general multi-task paradigm that combines shared feature extraction networks with bidirectional interactions. This paradigm models the associations between different traits indiscriminately: on the one hand, redundant information in auxiliary traits that is irrelevant to the target trait interferes with the feature learning of the target trait; on the other hand, bidirectional noise propagation between the target trait and auxiliary traits can mask the truly valuable complementary information between traits, resulting in the inability to accurately capture effective genetic associations, ultimately limiting the improvement of prediction accuracy.
[0004] To address this challenge, this invention proposes a multi-trait genome-wide prediction method based on unidirectional cross-trait attention. This method constructs a unidirectional asymmetric attention mechanism to directionally mine effective genetic associations between auxiliary traits and the target trait, achieving precise enhancement of the target trait's feature information by the auxiliary trait. This method fundamentally avoids the reverse interference caused by traditional bidirectional interactions, thereby significantly improving the prediction accuracy of multi-trait genome-wide predictions (GS). Summary of the Invention
[0005] The purpose of this invention is to provide a multi-trait genome-wide prediction method based on unidirectional cross-trait attention to address the following issues: Existing deep learning MT models mainly adopt the traditional multi-task learning paradigm, implicitly modeling the correlation between traits only through shared feature extraction networks, while lacking explicit information interaction mechanisms between different trait tasks. This symmetrical and equal interaction method has the following problems: First, it is difficult to achieve targeted information enhancement of the target trait by auxiliary traits, thus limiting prediction accuracy; second, redundant information between different traits may interfere with each other, introducing noise and affecting the robustness of the model.
[0006] To achieve the above objectives, the present invention provides the following solution: A multi-trait genome-wide prediction method based on unidirectional cross-trait attention includes: Obtain the genotype data of the individual to be predicted; The genotype data is input into a pre-trained multi-trait prediction model; the multi-trait prediction model includes a two-branch feature extraction module, a one-way cross-trait attention enhancement module, and a prediction module; The dual-branch feature extraction module is used to extract target trait features and at least one auxiliary trait feature based on the genotype data through a first branch and a second branch in parallel. The unidirectional cross-trait attention enhancement module is used to use the target trait feature as the query and the auxiliary trait feature as the key and value, calculate the cross-trait attention weight through a scaling dot product attention mechanism, perform weighted aggregation on the auxiliary trait feature to obtain aggregated features, and fuse the aggregated features with the target trait feature through a gating mechanism and residual connection to obtain the enhanced target trait feature. The prediction module is used to obtain the target trait prediction result based at least on the enhanced target trait features.
[0007] Preferably, in the unidirectional cross-trait attention enhancement module, the target trait feature is used as the query, and the auxiliary trait feature is used as the key and value. Cross-trait attention weights are calculated using a scaling dot product attention mechanism, and the auxiliary trait features are weighted and aggregated to obtain aggregated features. Specifically, this includes: A query matrix is generated by performing a linear transformation on the target trait features, and a key matrix and a value matrix are generated by performing a linear transformation on the auxiliary trait features; The scaled dot product of the query matrix and the key matrix is calculated, and the cross-trait attention weights are obtained after Softmax normalization. The aggregated features are obtained by weighting and aggregating the value matrix based on the cross-trait attention weights.
[0008] Preferably, in the unidirectional cross-trait attention enhancement module, the step of fusing the aggregated features with the target trait features through a gating mechanism and residual connections to obtain the enhanced target trait features specifically includes: The target trait features are subjected to linear transformation and Sigmoid activation to obtain gating weights; Calculate the Hadamard product of the gate weights and the aggregated features to obtain the gated aggregated features; The gated aggregated features and the target trait features are added together by residual addition to obtain the enhanced target trait features.
[0009] Preferably, in the dual-branch feature extraction module, the first branch and the second branch are both multilayer perceptrons, and each multilayer perceptron includes a first fully connected layer, a first activation function layer, a regularization layer, a second fully connected layer, and a second activation function layer connected in sequence.
[0010] Preferably, the prediction module includes a first prediction branch and a second prediction branch in parallel; the first prediction branch is used to output the predicted value of the target trait based on the enhanced target trait features, and the second prediction branch is used to output the predicted value of the auxiliary trait based on the auxiliary trait features.
[0011] Preferably, the training process of the multi-trait prediction model includes: The losses of each trait were normalized. Based on the homoscedastic uncertainty weighting, the normalized loss of each trait is dynamically fused to obtain the total loss; The model parameters are updated based on the total loss.
[0012] Preferably, the training process further includes: During training, training is terminated when the correlation coefficient between the true and predicted values of the target trait in the test set does not improve within a preset number of consecutive training rounds.
[0013] Preferably, the genotype data is single nucleotide polymorphism (SNP) site data obtained by filtering out invalid sites with a minimum allele frequency of less than 0.05 and a marker deletion rate of greater than 0.2; the target trait is an agronomic trait, and the auxiliary traits include agronomic traits and / or spectral traits.
[0014] Compared with the prior art, the present invention has the following advantages and technical effects: This invention abandons the general multi-task paradigm in deep learning MT models that relies solely on shared feature extraction networks and indiscriminately models trait associations. By constructing a unidirectional cross-trait attention mechanism, it achieves directional information enhancement from auxiliary traits to target traits, effectively avoiding redundant noise interference caused by bidirectional interaction in traditional multi-task learning, thereby significantly improving the GS prediction accuracy of target traits.
[0015] The method of this invention was used to predict four target traits of maize inbred lines: plant height, ear height, ear weight, and ear grain weight. For each target trait, three other agronomic traits and four raw spectra (red, green, near-infrared, and red-edge) were selected as auxiliary traits, and the performance of UniTANet was evaluated through cross-validation. The results are as follows: Figure 2As shown, UniTANet exhibits significant improvements in predictive power across all four traits compared to the single-trait GBLUP model. Specifically: when one of the remaining three agronomic traits is used as an auxiliary trait, the predictive power for plant height, ear height, ear weight, and ear grain weight increases by an average of 12.84%, 21.39%, 17.82%, and 15.63%, respectively; when one of the four original spectra is used as an auxiliary trait, the predictive power increases by an average of 13.58%, 19.78%, 17.17%, and 16.30%, respectively; when all three remaining agronomic traits are used as auxiliary traits, the predictive power increases by 14.83%, 19.66%, 17.16%, and 15.68%, respectively; and when all four original spectra are used as auxiliary traits, the predictive power increases by 14.15%, 22.14%, 18.01%, and 15.18%, respectively. Meanwhile, compared with single-trait MLP models that only extract target features, UniTANet's predictive power improved by 3.94%~8.52%, 4.62%~7.77%, 5.77%~7.75%, and 5.15%~9.20% under the four auxiliary trait configurations mentioned above, showing a stable and significant advantage.
[0016] Furthermore, UniTANet's prediction performance significantly outperforms bidirectional cross-trait attention models, an advantage that is particularly pronounced when the auxiliary trait is the original spectrum (e.g., Figure 3 (As shown in the figure). Specifically, when the auxiliary trait is a single agronomic trait, UniTANet improves the predictive power of plant height and ear height by an average of 6.72% and 5.39%, respectively, while the predictive effects of ear weight and ear grain weight are similar to those of the two-way model. When three agronomic traits are used as auxiliary traits simultaneously, UniTANet improves the predictive power of plant height and ear height by 7.81% and 6.17%, respectively, while the predictive effects of ear weight and ear grain weight remain similar. When the auxiliary trait is a single original spectrum, UniTANet improves the predictive power of the four traits by an average of 74.64%, 68.43%, 47.13%, and 45.20%, respectively. When four original spectra are used as auxiliary traits simultaneously, the improvement in predictive power further increases, reaching 91.65%, 79.14%, 64.60%, and 53.36%, respectively.
[0017] In summary, the UniTANet framework proposed in this invention can significantly improve the prediction accuracy of multi-trait GS models for maize inbred line phenotypes, providing reliable technical support for the efficient screening of superior backbone inbred lines, and thus promoting the development of precision maize breeding. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a diagram of the UniTANet framework structure, where, Figure 1 In the diagram, A represents the structure of the dual-branch feature extraction module. Figure 1 B in the diagram is a schematic diagram of the unidirectional cross-trait attention enhancement module structure. Figure 1 C in the diagram is a schematic of the prediction head module structure; Figure 2 This is a graph showing the evaluation results of UniTANet's predictive power for four agronomic traits of maize inbred lines. Figure 2 In the diagram, A represents the predictive force box plot when plant height is the target trait. Figure 2 B in the figure represents the predictive force box plot when ear height is the target trait. Figure 2 C in the figure represents the predictive force box plot when ear weight is the target trait. Figure 2 D in the figure represents the predictive force box plot when the grain weight of the ear is taken as the target trait; Figure 3 The figure shows a comparison of the prediction performance of UniTANet and the bidirectional cross-trait attention model on four agronomic traits of maize inbred lines. Figure 3 In the figure, A is a bar chart comparing the predictive power when plant height is used as the target trait. Figure 3 B in the chart represents a comparison of predictive power when ear height is used as the target trait. Figure 3 In the figure, C is a bar chart comparing the predictive power when ear weight is used as the target trait. Figure 3 D in the figure is a bar chart comparing the predictive power when the grain weight of the ear is taken as the target trait. Detailed Implementation
[0019] 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 scope of protection of the present invention.
[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0021] Example 1 This embodiment fully demonstrates the specific implementation process and performance verification results of the multi-trait genome-wide prediction method based on unidirectional cross-trait attention on a maize inbred line dataset, including: I. Data Acquisition and Preprocessing Genomic and phenotypic data of maize inbred lines were obtained. Genomic SNPs were screened for quality control, and invalid sites with a minimum allele frequency of less than 0.05 and a marker deletion rate of greater than 0.2 were filtered out, ultimately retaining 99,976 valid SNPs. Phenotypic data included four agronomic traits: plant height (PH), ear height (EH), ear weight (EW), and ear grain weight (EGW), as well as four raw spectra: red light (R), green light (G), near-infrared (NIR), and red edge (RedE).
[0022] II. Setting of Target and Auxiliary Traits One of the four agronomic traits—plant height, ear height, ear weight, and ear grain weight—was selected as the target trait, while the remaining agronomic traits and the four original spectra were selected as candidate auxiliary traits. Two auxiliary trait configuration schemes were set up: (1) Single auxiliary trait scheme: only one of the remaining three agronomic traits or one of the four original spectra was selected as an auxiliary trait; (2) Multiple auxiliary trait scheme: the remaining three agronomic traits were selected as auxiliary traits at the same time, or the four original spectra were selected as auxiliary traits at the same time.
[0023] III. Dataset Partitioning The preprocessed multi-trait dataset was randomly divided into 10 parts using 10-fold cross-validation. One part was used as the test set, and the remaining 9 parts were used as the training set.
[0024] Model performance was evaluated using the Pearson correlation coefficient (PCC) between the true and predicted values of the target trait in the test set. The optimal PCC obtained during training was taken as the final single-fold result. After completing 10-fold testing, the average of the 10 PCCs was calculated as the predictive power of one cross-validation. To ensure the stability and reliability of the results, the cross-validation process was repeated 10 times, and the average of the 10 predictive powers was taken as the model evaluation result.
[0025] IV. Constructing a multi-trait genome-wide prediction framework (UniTANet) The constructed multi-trait genome-wide prediction framework UniTANet, such as Figure 1 As shown, it consists of three modules connected in series: a dual-branch feature extraction module, a unidirectional cross-trait attention enhancement module, and a prediction head module.
[0026] 1. Dual-branch feature extraction module: structure as follows Figure 1As shown in A, this module takes 99,976-dimensional genotype data as input and employs two identical parallel multilayer perceptron (MLP) branches, corresponding to the target trait branch and the auxiliary trait branch, respectively. Regardless of whether the input is a single auxiliary trait or a combination of multiple auxiliary traits, unified feature extraction is performed through the auxiliary trait branch. Each branch contains two fully connected layers: the first layer reduces the 99,976-dimensional input to a 256-dimensional feature space, performs a nonlinear transformation using the ReLU activation function, and then performs regularization through a Dropout layer (dropout rate of 0.1); the second layer further compresses the features to 64 dimensions and uses the ReLU activation function again to enhance the nonlinear expressive power. Finally, the two branches output 64-dimensional target features (Target_feat) and auxiliary features (Aux_feat), respectively, achieving efficient decoupled extraction from high-dimensional genotype data to low-dimensional structured features.
[0027] 2. Unidirectional transtrait attention enhancement module: such as Figure 1 As shown in B, this module employs a unidirectional cross-trait attention mechanism to achieve unidirectional information enhancement of the target trait by auxiliary traits. Specifically, the target feature is used as the query, and the auxiliary features are used as the key and value. Cross-trait attention weights are calculated through scaling dot products and Softmax normalization. Based on these weights, the auxiliary features are adaptively weighted and aggregated to obtain the attention-aggregated feature (Aux_context). The calculation process can be represented as follows: (1)
[0028] in, Q From target features Generated through linear transformation , K and V Auxiliary features Generated through linear transformation , , It represents the Hadamah accumulation. a for A column vector whose elements are all 1s. A This represents attention aggregation features. m The number of samples per batch (in this embodiment) ).
[0029] Furthermore, this module introduces a gating mechanism and residual connections to dynamically adjust the fusion strength of aggregated information while preserving the original target features, thereby selectively enhancing auxiliary trait information. The specific calculation process is as follows: (2)
[0030] in, This refers to the enhanced target trait feature (Enhanced_target_feat). W The weight matrix is a linear transformation matrix. b This is a bias term.
[0031] 3. Prediction Head Module: such as Figure 1 As shown in C, this is the output of the entire architecture. It adopts a parallel dual-branch output structure. One branch is the target trait prediction branch, which converts Enhanced target_feat into the target trait prediction result through a fully connected mapping. The other branch is the auxiliary trait prediction branch, which can flexibly adjust the output dimension according to the actual number of auxiliary traits, and simultaneously complete the parallel prediction of multiple auxiliary traits.
[0032] V. Model Training and Optimization To avoid bias in training caused by the magnitude of loss across different tasks, the loss for each trait is normalized. First, before training, the baseline loss for the target trait and each auxiliary trait is calculated for all batches in the training set, as follows: (3)
[0033] (4)
[0034] in, b It is the total number of batches in the training set. m The number of samples per batch (in this embodiment) ), It is the first k The first batch j The true value of the target trait for each sample It is the first k The first batch j Predicted values of the target trait for each sample It is the first k The first batch i The first of the auxiliary traits j The true value of each sample It is the first k The first batch i The first of the auxiliary traits j The predicted value for each sample, with the baseline value remaining unchanged throughout the training iterations.
[0035] During training, the losses for the target trait and auxiliary traits in the current batch are first calculated and normalized. The specific formula is as follows: (5)
[0036] (6)
[0037] in, and These are the target trait and the first i The first of the auxiliary traits k Losses in each batch and It is the loss after normalization.
[0038] To further achieve a dynamic balance of losses across multiple traits, we dynamically fuse the normalized losses of each trait based on homoscedastic uncertainty weighting. The total loss is calculated as follows: (7)
[0039] in, It is the first k Total loss under the batch, and , and They are the first k Loss weights for target and auxiliary traits in batches. and It is a penalty item. and These are learnable parameters, and in this embodiment, the initial values are 0.6 and 0.4, respectively.
[0040] Finally, backpropagation and parameter updates are performed based on the total loss. Furthermore, to avoid overfitting, an early stopping mechanism is used: model training is stopped when the PCC (Plan-C between the true and predicted values of the target trait in the test set) no longer increases for 50 consecutive training epochs.
[0041] VI. Model Performance Validation To verify the technical effectiveness of this invention, the schemes described in (I) to (V) above were systematically validated on a maize inbred line dataset. The comparison models include: 1. Unitrait GBLUP model (traditional statistical genetics baseline) 2. Single-trait MLP model (containing only the target feature extraction branch, without auxiliary trait branches and attention enhancement module) 3. Bidirectional cross-trait attention model (which adopts the same dual-branch MLP backbone as UniTANet, but in the attention module, the target feature and auxiliary feature are simultaneously used as query, key and value for each other, and bidirectional attention is calculated) The verification results are as follows: (1) Compared with the single-trait GBLUP model, UniTANet significantly improved the predictive power of all four traits. Specifically: when one of the remaining three agronomic traits was used as an auxiliary trait, the predictive power of plant height, ear height, ear weight, and ear grain weight increased by an average of 12.84%, 21.39%, 17.82%, and 15.63%, respectively; when one of the four original spectra was used as an auxiliary trait, the predictive power increased by an average of 13.58%, 19.78%, 17.17%, and 16.30%, respectively; when the remaining three agronomic traits were used as auxiliary traits, the predictive power increased by 14.83%, 19.66%, 17.16%, and 15.68%, respectively; and when all four original spectra were used as auxiliary traits, the predictive power increased by 14.15%, 22.14%, 18.01%, and 15.18%, respectively.
[0042] (2) Compared with the single-trait MLP model, UniTANet’s predictive power improved by 3.94%~8.52%, 4.62%~7.77%, 5.77%~7.75% and 5.15%~9.20% under the above four auxiliary trait configurations.
[0043] (3) Compared with the two-way cross-trait attention model, UniTANet has significantly better predictive performance, which is particularly prominent when the auxiliary trait is the original spectrum. When the auxiliary trait is a single agronomic trait, UniTANet improves the predictive power of plant height and ear height by an average of 6.72% and 5.39%, respectively; when three agronomic traits are used as auxiliary traits at the same time, the predictive power of plant height and ear height is improved by 7.81% and 6.17%, respectively; when the auxiliary trait is a single original spectrum, the predictive power of the four traits is improved by an average of 74.64%, 68.43%, 47.13%, and 45.20%, respectively; when four original spectra are used as auxiliary traits at the same time, the improvement in predictive power is further increased, reaching 91.65%, 79.14%, 64.60%, and 53.36%, respectively.
[0044] The above verification results are attached. Figure 2 and attached Figure 3 As shown: Figure 2 The predictive power of UniTANet for four agronomic traits of maize inbred lines is presented in the form of box plots. Figure 2In the diagram, subplots A, B, C, and D use plant height, ear height, ear weight, and ear grain weight as target traits, respectively. The horizontal axis of each subplot represents different auxiliary trait configurations (including single agronomic trait, single raw spectrum, multiple agronomic traits, and multiple raw spectra), while the vertical axis represents predictive power (Pearson correlation coefficient). Each box plot, from left to right, shows the predictive power distribution of three methods: single-trait GBLUP, single-trait MLP, and the UniTANet method of this invention. Different lowercase letters in the figures indicate significant differences between the methods.
[0045] Figure 3 The prediction performance of the present invention UniTANet and the bidirectional cross-trait attention model is shown in the form of a bar chart. Figure 3 In the diagram, subplots A, B, C, and D use plant height, ear height, ear weight, and ear grain weight as target traits, respectively. The horizontal axis of each subplot represents different auxiliary trait configurations, and the vertical axis represents predictive power. The left bar in each subplot represents the mean predictive power of the UniTANet model, and the right bar represents the mean predictive power of the two-way cross-trait attention model. The error bars represent the standard deviation of multiple repetitions, and "**" in the diagram indicates a significant difference between the two models.
[0046] Therefore, the UniTANet framework proposed in this invention can significantly improve the prediction accuracy of multi-trait GS models for maize inbred line phenotypes, providing reliable technical support for efficient screening of superior backbone inbred lines, and thus promoting the development of precision maize breeding.
[0047] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A multi-trait genome-wide prediction method based on unidirectional cross-trait attention, characterized in that, include: Obtain the genotype data of the individual to be predicted; The genotype data is input into a pre-trained multi-trait prediction model; the multi-trait prediction model includes a two-branch feature extraction module, a one-way cross-trait attention enhancement module, and a prediction module; The dual-branch feature extraction module is used to extract target trait features and at least one auxiliary trait feature based on the genotype data through a first branch and a second branch in parallel. The unidirectional cross-trait attention enhancement module is used to use the target trait feature as the query and the auxiliary trait feature as the key and value, calculate the cross-trait attention weight through a scaling dot product attention mechanism, perform weighted aggregation on the auxiliary trait feature to obtain aggregated features, and fuse the aggregated features with the target trait feature through a gating mechanism and residual connection to obtain the enhanced target trait feature. The prediction module is used to obtain the target trait prediction result based at least on the enhanced target trait features.
2. The multi-trait genome-wide prediction method based on unidirectional cross-trait attention as described in claim 1, characterized in that: In the unidirectional cross-trait attention enhancement module, the target trait feature is used as the query, and the auxiliary trait feature is used as the key and value. Cross-trait attention weights are calculated using a scaled dot product attention mechanism, and the auxiliary trait features are weighted and aggregated to obtain aggregated features. Specifically, this includes: A query matrix is generated by performing a linear transformation on the target trait features, and a key matrix and a value matrix are generated by performing a linear transformation on the auxiliary trait features; The scaled dot product of the query matrix and the key matrix is calculated, and the cross-trait attention weights are obtained after Softmax normalization. The aggregated features are obtained by weighting and aggregating the value matrix based on the cross-trait attention weights.
3. The multi-trait genome-wide prediction method based on unidirectional cross-trait attention as described in claim 1, characterized in that: In the unidirectional cross-trait attention enhancement module, the step of fusing the aggregated features with the target trait features through a gating mechanism and residual connections to obtain the enhanced target trait features specifically includes: The target trait features are subjected to linear transformation and Sigmoid activation to obtain gating weights; Calculate the Hadamard product of the gate weights and the aggregated features to obtain the gated aggregated features; The gated aggregated features and the target trait features are added together by residual addition to obtain the enhanced target trait features.
4. The multi-trait genome-wide prediction method based on unidirectional cross-trait attention as described in claim 1, characterized in that: In the dual-branch feature extraction module, the first branch and the second branch are both multilayer perceptrons, and each multilayer perceptron includes a first fully connected layer, a first activation function layer, a regularization layer, a second fully connected layer, and a second activation function layer connected in sequence.
5. The multi-trait genome-wide prediction method based on unidirectional cross-trait attention according to claim 1, characterized in that: The prediction module includes a first prediction branch and a second prediction branch in parallel; the first prediction branch is used to output the predicted value of the target trait based on the enhanced target trait features, and the second prediction branch is used to output the predicted value of the auxiliary trait based on the auxiliary trait features.
6. A multi-trait genome-wide prediction method based on unidirectional transtrait attention according to any one of claims 1-5, characterized in that: The training process of the multi-trait prediction model includes: The losses of each trait were normalized. Based on the homoscedastic uncertainty weighting, the normalized loss of each trait is dynamically fused to obtain the total loss; The model parameters are updated based on the total loss.
7. The multi-trait genome-wide prediction method based on unidirectional cross-trait attention according to claim 6, characterized in that: The training process also includes: During training, training is terminated when the correlation coefficient between the true and predicted values of the target trait in the test set does not improve within a preset number of consecutive training rounds.
8. A multi-trait genome-wide prediction method based on unidirectional cross-trait attention according to any one of claims 1-5, characterized in that: The genotype data is single nucleotide polymorphism (SNP) site data obtained by filtering out invalid sites with a minimum allele frequency of less than 0.05 and a marker deletion rate of greater than 0.2; the target trait is an agronomic trait, and the auxiliary traits include agronomic traits and / or spectral traits.