Method for constructing a deep learning-based prediction model for salt-tolerant traits of brassica campestris

Abstract

The application discloses a method for constructing a rapeseed salt-tolerant trait prediction model based on deep learning, comprising collecting multi-modal data of a preset rapeseed area, and preprocessing the multi-modal data; constructing a three-modal coupling encoder with biological prior constraints according to the multi-modal data, and constructing a cross-species migration mechanism based on gene homologous mapping to obtain a mapping library for a rapeseed small sample scene; constructing a strategy optimization mechanism with cooperation of outer meta-learning and inner reinforcement learning, and constructing a rapeseed salt-tolerant trait prediction model according to the three-modal coupling encoder, the mapping library and the strategy optimization mechanism; inputting to-be-predicted data into the rapeseed salt-tolerant trait prediction model to obtain prediction data, obtaining a prediction variance according to the prediction data, triggering an active learning sampling mechanism when the prediction variance is greater than a variance threshold, and adding a current sample into a training support set to output a target model.

Description

Technical Field

[0001] This invention relates to the field of rapeseed salt-alkali tolerance trait prediction technology, and in particular to a method for constructing a rapeseed salt-alkali tolerance trait prediction model based on deep learning. Background Technology

[0002] Severe soil salinization severely restricts rapeseed yield and quality, and cultivating salt-tolerant varieties is a key approach to improving and utilizing saline-alkali land. Currently, the identification of rapeseed salt-alkali tolerance traits relies mainly on field phenotypic observations and physiological index measurements, which suffer from problems such as long cycles, low efficiency, and susceptibility to environmental interference, making it difficult to meet the needs of large-scale precision breeding.

[0003] With the development of multi-omics technologies, multimodal data such as genotypes, phenotypes, and environments are rapidly accumulating. However, traditional prediction methods struggle to efficiently integrate high-dimensional heterogeneous information and are insufficient in characterizing the nonlinear relationships between genes and between genes and the environment. Furthermore, labeled samples related to rapeseed salt tolerance are scarce, limiting the generalization ability of existing models. These models also lack biological prior constraints, resulting in poor interpretability and an inability to effectively integrate with biological mechanisms such as salt tolerance gene pathways.

[0004] Furthermore, the rigid model training strategies make it difficult to dynamically adapt to scenarios such as small samples and class imbalance, and their insufficient active learning and knowledge transfer capabilities result in trait prediction accuracy and reliability that cannot support actual breeding decisions. Existing technologies cannot achieve multimodal information collaborative representation, cross-species knowledge transfer, and adaptive strategy optimization. There is an urgent need to construct novel prediction methods that integrate biological mechanisms and deep learning to provide technical support for precision breeding of salt-alkali tolerant rapeseed. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing a prediction model of rapeseed salt tolerance based on deep learning.

[0006] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: Multimodal data of a predefined rapeseed region is collected and preprocessed; the multimodal data includes phenotypic modality data, genotypic modality data, and environmental modality data. Based on the multimodal data, a three-modal coupled encoder with biological prior constraints is constructed. For the small sample scenario of rapeseed, a cross-species migration mechanism based on gene homology mapping is constructed to obtain a mapping library. A policy optimization mechanism that combines outer-layer meta-learning and inner-layer reinforcement learning is constructed, and a rapeseed salt-alkali tolerance trait prediction model is built based on the three-modal coupled encoder, the mapping library, and the policy optimization mechanism. The data to be predicted is input into the rapeseed salt-alkali tolerance trait prediction model to obtain prediction data. The prediction variance is obtained based on the prediction data. When the prediction variance is greater than the variance threshold, the active learning sampling mechanism is triggered to include the current sample into the training support set and output the target model.

[0007] Furthermore, the method for constructing a biologically prior-constrained trimodal coupled encoder based on the multimodal data includes: Obtain the phenotypic tensor, genotype matrix, and environment tensor, and construct a three-modal coupled encoder using phenotypic coding branch, genotype coding branch, and environment coding branch; Phenotypic encoding branch: Employs a lightweight multispectral vision Transformer, inputting phenotypic tensors. By employing patch embedding, reverse residual bottleneck, and global attention mechanisms, phenotypic latent vectors are output. Genotype coding branch: A graph attention network is constructed based on the gene interaction network, where nodes are single nucleotide polymorphism sites and edges represent linkage disequilibrium. The input is the genotype matrix. Output genotype latent vector Environment encoding branch: Employs a temporal convolutional network to encode the dynamics of the environment, with the input environment tensor... Output environment latent vector ; Constructing a semantic anchor matrix for salt and alkali tolerance gene pathways ,in For the prior number of paths, As a dimension, each row Let p be the semantic center vector of the p-th pathway; define the biological semantic contrast loss. : ; in For varieties The set of annotated salt and alkali tolerance pathways, For a set of irrelevant paths, For the interval hyperparameter, For variety index, For varieties Genotype latent vector, For set The number of intermediate pathways, The square of the L2 norm of the vector; The phenotypic latent vector, genotype latent vector, and environmental latent vector are fused into a joint representation through a cross-modal attention fusion layer, expressed as: ; ; in For queryable vectors, For modal adaptive weights, For joint representation vector, For modal indexing, Let be the learnable weight matrix for the m-th mode. This is the transpose of the queryable vector.

[0008] Furthermore, methods for constructing a cross-species migration mechanism mapping library based on gene homology mapping include: We collected whole-genome protein-coding sequences and functional annotation data of source and target species, and performed low-complexity region filtering and signal peptide splicing preprocessing on the protein sequence sets of source and target species to obtain a standardized genome dataset. A dynamic programming algorithm is used to perform full-pair whole-sequence alignment between the source and target species to construct an initial similarity score matrix. The Smith-Waltman algorithm is then used to calculate the global alignment score between any source and target sequences. ;in To compare the local optima of the score matrix, For varieties The source sequence, For variety j, For the first Linear vacancy penalty for rows, For the first Linear empty space penalty for columns, To compare the score matrix Line number Column elements, Score for amino acid residue substitution. And when Forced reset to zero; For each pair Perform bidirectional optimal comparison and screening: forward reverse Candidate homologous pairs are retained when both bidirectional optimal conditions are met. The scores of candidate homologous pairs are normalized; where A collection of protein sequences from the source species. A set of protein sequences of the target species; Construct a weighted homology graph Vertex set edge set Edge weight matrix If an edge exists, then the element Otherwise, it is 0; where For similarity threshold, For varieties The edge weight with respect to variety j; Markov clustering algorithm is used to analyze the weighted homology graph. Homologous gene clusters were divided, and a transition probability matrix was constructed. Execute extension With expansion The process stops when the entropy change of the transition probability matrix is ​​less than the entropy change threshold, thus obtaining a set of homologous gene clusters; where... For expansion parameters, Let x be the probability of moving from a random walk to a node h. Calculate homologous gene clusters Intraspecific target species genes Genes of endogenous species in the collection of homologous gene clusters direct homology confidence between The orthologous homology confidence scores within each cluster are aggregated into a global gene homology mapping matrix. ;in ; This represents the homology confidence between the target gene and the source species gene. For evolutionary conservation function, The attenuation coefficient is... For the target sequence Number of orthologous genes in other sequenced species. To compare advantages, Penalty for the number of endogenous genes within homologous gene clusters; Only source gene mappings with a confidence level greater than a confidence threshold are retained, and Softmax normalization of the global gene homology mapping matrix is ​​performed along the source gene dimension. For gene clusters with many-to-many mappings, a chromosomal neighborhood collinearity constraint is introduced to obtain the mapping matrix. ;in Based on gene neighborhood collinearity scores, For temperature parameters; The mapping matrix after quality control The metadata of homologous gene clusters is encapsulated into a structured mapping library. ;in A set of gene function annotation vectors for the target species. This is the set of pre-trained representation vectors of the source species' genes. To divide homologous gene clusters into sets, It is a collection of homologous gene clusters.

[0009] Furthermore, methods for constructing a policy optimization mechanism that coordinates outer-layer meta-learning and inner-layer reinforcement learning include: Outer Meta-Learning: Employing gradient-based efficient meta-learning on a task distribution composed of multiple salt tolerance trait-related tasks. Learning model initial parameters ,in To execute on the k-th task Task-specific parameters after gradient descent. The meta-learning rate, The total number of tasks. This represents the number of gradient descent steps in the inner loop. These are the initial parameters; Inner reinforcement learning: Each training batch is treated as a Markov decision process, and the training strategy is dynamically adjusted using a proximal policy optimization algorithm; where the state space... The joint characterization of the distribution entropy of the current batch of samples Conflict angles of gradients for each mode of loss Validation set performance degradation rate and sample class imbalance The components are assembled, and the dimensions of the assembly are unified. Action space Including the sampling probability adjustment vector for each modality Loss weighting coefficient and learning rate scaling factor Policy Network: A two-layer MLP is used to approximate the policy function. and value function Reward function Improved accuracy on the comprehensive validation set The sample efficiency gain and training stability are expressed as follows: ;in For classification loss weights, For biological semantic contrast loss weights, For L2 regularization loss weights, For the policy function, Let v be the reward value for the v-th training step. This represents the number of labeled samples in the current batch. These are the weighting coefficients for the sample efficiency term. To train the weight coefficients of the stability term, The variance of the training loss; At each training step, the policy network outputs an action. Dynamic adjustment: based on the status The sampling probabilities of each modality are reweighted, prioritizing the sampling of difficult samples with high information gain, by adjusting... Balance the classification loss, biological semantic contrast loss, and L2 regularization loss, and dynamically scale the learning rate according to the training phase; The policy network is trained by optimizing the pruning objective function using a proximal strategy. The expression is: ; in For probability ratios, For generalized advantage estimation, The objective function for optimizing the near-end strategy is... For the expected value of all training steps, To truncate hyperparameters, This is a truncation function.

[0010] Furthermore, the method for constructing the rapeseed salt-alkali tolerance prediction model includes: Using 7-tuple A predictive model for salt and alkali tolerance traits in rapeseed was developed, and differential parameter initialization was performed on each component of the seven-tuple; among which... It is a stereo encoder. It is a genotype encoder. For environmental encoders, For cross-modal attention fusion module, For cross-species migration module, For policy networks, For prediction head; The rapeseed salt-alkali tolerance prediction model includes encoder parameter initialization and prior injection, cross-species migration mechanism embedded encoder, and meta-reinforcement collaborative training algorithm; Encoder parameter initialization and prior injection: semantic anchor matrix of salt tolerance gene pathway The embedded genotype encoder optimizes the objective by updating the semantic anchor matrix of salt tolerance gene pathways via exponential moving average during training. ,in For the first The semantic anchor vector of the p-th path during training, The momentum coefficient, For the first The semantic anchor vector of the p-th path during training, For the first The subset of samples annotated to the p-th path during training. For the first The genotype feature vector of each gene sample output by the model in the current round of training; Cross-species migration mechanisms embedded in encoders: in the genotype coding branch After the graph attention layer outputs, a cross-species transfer module is embedded. : Hidden features of the target gene are mapped using a library Retrieval of source species pre-trained representations Calculate the gate vector and output the transfer enhancement feature. ;in For the target species after interspecies migration The hidden layer feature vectors of each gene in the encoder output. For the target species The first gene and its source species The reliability of homology relationships among individual genes A learnable transfer matrix, For the source species Pre-trained representation of each gene, This represents the number of homologous genes retrieved in the source species. It is the sigmoid activation function. The learnable gated weight matrix, For adaptive gating vectors, For element-wise multiplication, For the target species The hidden layer feature vectors of each gene in the encoder output; Meta-reinforcement collaborative training algorithm: Defining the task distribution for predicting rapeseed salt and alkali tolerance traits Each task corresponds to a specific salinization stress condition, and each task includes a support set and a query set. The outer loop objective of the meta-learning is to find the optimal initial parameters, and a first-order approximation model-independent meta-learning is adopted. ;in These are the initial parameters. For the model in the task The parameters on the support set after n steps of gradient descent The total number of tasks. The meta-learning rate; Each training batch is treated as a time step in a Markov decision process. An inner loop for policy optimization is constructed, and a state vector is built at each training step. An action-space policy network is defined. Output Action ;in Let v be the state vector at the v-th training step. For the action of the v-th training step, Let v be the learning rate scaling factor for the v-th training step. Assign coefficients to the loss weights for the v-th training step. This is the sampling probability adjustment vector for the v-th training step; Given a reward function The network parameters of the update strategy are optimized using a near-end strategy, based on task distribution. Randomly sample K tasks, and for each task Execute inner layer adaptation, during policy network Each step outputs an action. Adjust training hyperparameters and evaluate adapted parameters on the query set. Calculate task loss and reward, and aggregate all task gradients to update the initial parameters. Network parameters based on cumulative trajectory update strategy ; Predicting Head A three-layer fully connected network is adopted, with the hidden layer dimensions being as follows: BatchNorm and LeakyReLU activations are inserted between layers, and the output layer uses Sigmoid activation to constrain the predicted values ​​to [0,1]. The expression is as follows: ;in Here are the learnable parameters of the prediction head; where The dimension of the joint representation of z after cross-modal fusion; The composite loss function is composed of three sub-losses dynamically weighted by the policy network. ,in For classifying losses, For biological semantic contrast loss, For parameter regularization loss, For classification weights, For biological semantic contrast weights, Regularize the weights for the parameters; Training is stopped and the target model is output when the mean absolute error of the validation set does not decrease for 10 consecutive meta-iterations.

[0011] Furthermore, the method for obtaining the prediction variance based on the prediction data includes: Multi-source heterogeneous data of the rapeseed variety to be predicted are input into the trained three-modal coupled encoder to obtain joint representations. Input prediction header: ;in The comprehensive index of salt and alkali tolerance. It is a multilayer perceptron; Forward propagation using the Monte Carlo dropout method Next, calculate the prediction variance. : ; ; in To predict the mean, As a priori for model accuracy, Forward propagation count, is the comprehensive index of salt and alkali tolerance during the t-th forward propagation.

[0012] Furthermore, the expression for the variance threshold is: ; in Let be the variance threshold at the t-th forward propagation. This is the global conservative coefficient. To validate the q-quantile function of the variance distribution predicted by the validation set, The set of predicted variances for the validation set samples. This represents the relaxation factor during early training. The dynamic decay rate is t, where t is the current training round. To adjust the strength for category scarcity, It is the numerical stability constant. For varieties Category The current number of labeled samples.

[0013] The beneficial effects of this invention are: This invention relates to a method for constructing a prediction model for salt-alkali tolerance traits in rapeseed based on deep learning. Compared with existing technologies, this invention has the following technical advantages: This invention constructs a biological prior-constrained three-modal coupled encoder that integrates phenotypic, genotypic, and environmental information, combined with gene pathway semantic constraints, to improve feature interpretability and prediction accuracy. It designs a gene homology cross-species migration mechanism to address the scarcity of rapeseed salt-alkali tolerant samples, utilizing homology knowledge to enhance feature expression and significantly improve generalization ability in small-sample scenarios. It employs outer-layer meta-learning and inner-layer reinforcement learning for collaborative optimization, adaptively adjusting the training strategy, balancing multiple loss weights and learning rate, resulting in more stable and efficient training. Furthermore, it introduces prediction variance-driven active learning to automatically select high-value samples to expand the training set, continuously iterating and optimizing the model to improve prediction reliability and practicality. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the steps of constructing a deep learning-based prediction model for salt and alkali tolerance traits in rapeseed according to the present invention. Detailed Implementation

[0015] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0016] The method for constructing a deep learning-based prediction model for rapeseed salt-alkali tolerance in this invention includes the following steps: like Figure 1 As shown, this embodiment includes the following steps: Multimodal data of a predefined rapeseed region is collected and preprocessed; the multimodal data includes phenotypic modality data, genotypic modality data, and environmental modality data. Preprocessing includes: radiometric correction, band selection, and normalization of hyperspectral images, outputting phenotypic tensors; imputation of missing values, filtering of minor allele frequencies, and normalization encoding of single nucleotide polymorphism data, outputting genotype matrices; Z-score normalization and sliding window reconstruction of environmental time-series data, outputting environmental tensors; band selection: retaining the first k bands with information weights based on the ReliefF algorithm; imputation of missing values: kNN imputation based on linkage disequilibrium. In the actual assessment, the saline-alkali cultivated land in region U11 was taken as the research object, with soil salt content of 0.3%~0.6% and pH value of 8.2~8.8, which is a typical moderate saline-alkali stress environment; the experimental materials were 120 rapeseed varieties, including 35 salt-alkali tolerant varieties, 45 moderately salt-alkali tolerant varieties, and 40 salt-alkali sensitive varieties. Phenotypic modal data included plant height, fresh weight, root length, chlorophyll content, and leaf salt damage index; genotypic modal data included approximately 52,000 SNP loci in the rapeseed genome; and environmental modal data included average daily temperature, soil salinity, pH value, rainfall, and duration of sunlight. Based on the multimodal data, a three-modal coupled encoder with biological prior constraints is constructed. For the small sample scenario of rapeseed, a cross-species migration mechanism based on gene homology mapping is constructed to obtain a mapping library. A policy optimization mechanism that combines outer-layer meta-learning and inner-layer reinforcement learning is constructed, and a rapeseed salt-alkali tolerance trait prediction model is built based on the three-modal coupled encoder, the mapping library, and the policy optimization mechanism. The data to be predicted is input into the rapeseed salt-alkali tolerance trait prediction model to obtain prediction data. The prediction variance is obtained based on the prediction data. When the prediction variance is greater than the variance threshold, the active learning sampling mechanism is triggered to include the current sample into the training support set and output the target model.

[0017] In this embodiment, the method for constructing a biologically prior-constrained trimodal coupled encoder based on the multimodal data includes: Obtain the phenotypic tensor, genotype matrix, and environment tensor, and construct a three-modal coupled encoder using phenotypic coding branch, genotype coding branch, and environment coding branch; Phenotypic encoding branch: Employs a lightweight multispectral vision Transformer, inputting phenotypic tensors. By employing patch embedding, reverse residual bottleneck, and global attention mechanisms, phenotypic latent vectors are output. Genotype coding branch: A graph attention network is constructed based on the gene interaction network, where nodes are single nucleotide polymorphism sites and edges represent linkage disequilibrium. The input is the genotype matrix. Output genotype latent vector Environment encoding branch: Employs a temporal convolutional network to encode the dynamics of the environment, with the input environment tensor... Output environment latent vector ; Constructing a semantic anchor matrix for salt and alkali tolerance gene pathways ,in For the prior number of paths, As a dimension, each row Let p be the semantic center vector of the p-th pathway; define the biological semantic contrast loss. : ; in For varieties The set of annotated salt and alkali tolerance pathways, For a set of irrelevant paths, For the interval hyperparameter, For variety index, For varieties Genotype latent vector, For set The number of intermediate pathways, The square of the L2 norm of the vector; The phenotypic latent vector, genotype latent vector, and environmental latent vector are fused into a joint representation through a cross-modal attention fusion layer, expressed as: ; ; in For queryable vectors, For modal adaptive weights, For joint representation vector, For modal indexing, Let be the learnable weight matrix for the m-th mode. This is the transpose of the queryable vector; In practical evaluation, biological prior constraints are used to construct semantic anchors based on functional annotations of the SOS signaling pathway, NHX family, and HKT family. Phenotypic encoding branch: Lightweight multispectral ViT, patch size 16×16, 3 inverse residual bottleneck layers, 4 global attention heads, output phenotypic latent vector dimension 256; Genotypic encoding branch: Graph attention network, nodes are SNP sites, edges are gene linkage disequilibrium, 3 attention heads, output genotypic latent vector dimension 256; Environment encoding branch: Temporal convolutional network (TCN), dilation factor [1,2,4], convolution kernel size 3, output environment latent vector dimension 256. The core pathways for salt and alkali tolerance in rapeseed were selected, namely the SOS signaling pathway, the NHX ion transport pathway, and the HKT potassium-sodium regulatory pathway. A 3×256 semantic anchor matrix of salt and alkali tolerance gene pathways was constructed. Define bio-semantic contrast loss The interval hyperparameter is 0.8, which constrains the aggregation of genotype features toward the semantic center of the salt tolerance pathway; Modality-adaptive weights By fusing the three-modal latent vectors, a 768-dimensional joint representation is output. Automatically weighted high-contribution modes, among which It is 0.42. It is 0.35. It is 0.23.

[0018] In this embodiment, the method for constructing a cross-species migration mechanism mapping library based on gene homology mapping includes: We collected whole-genome protein-coding sequences and functional annotation data of source and target species, and performed low-complexity region filtering and signal peptide splicing preprocessing on the protein sequence sets of source and target species to obtain a standardized genome dataset. A dynamic programming algorithm is used to perform full-pair whole-sequence alignment between the source and target species to construct an initial similarity score matrix. The Smith-Waltman algorithm is then used to calculate the global alignment score between any source and target sequences. ;in To compare the local optima of the score matrix, For varieties The source sequence, For variety j, For the first Linear vacancy penalty for rows, For the first Linear empty space penalty for columns, To compare the score matrix Line number Column elements, Score for amino acid residue substitution. And when Forced reset to zero; For each pair Perform bidirectional optimal comparison and screening: forward reverse Candidate homologous pairs are retained when both bidirectional optimal conditions are met. The scores of candidate homologous pairs are normalized; where A collection of protein sequences from the source species. A set of protein sequences of the target species; Construct a weighted homology graph Vertex set edge set Edge weight matrix If an edge exists, then the element Otherwise, it is 0; where For similarity threshold, For varieties The edge weight with respect to variety j; Markov clustering algorithm is used to analyze the weighted homology graph. Homologous gene clusters were divided, and a transition probability matrix was constructed. Execute extension With expansion The process stops when the entropy change of the transition probability matrix is ​​less than the entropy change threshold, thus obtaining a set of homologous gene clusters; where... For expansion parameters, Let x be the probability of moving from a random walk to a node h. Calculate homologous gene clusters Intraspecific target species genes Genes of endogenous species in the collection of homologous gene clusters direct homology confidence between The orthologous homology confidence scores within each cluster are aggregated into a global gene homology mapping matrix. ;in ; This represents the homology confidence between the target gene and the source species gene. For evolutionary conservation function, The attenuation coefficient is... For the target sequence Number of orthologous genes in other sequenced species. To compare advantages, Penalty for the number of endogenous genes within homologous gene clusters; Only source gene mappings with a confidence level greater than a confidence threshold are retained, and Softmax normalization of the global gene homology mapping matrix is ​​performed along the source gene dimension. For gene clusters with many-to-many mappings, a chromosomal neighborhood collinearity constraint is introduced to obtain the mapping matrix. ;in Based on gene neighborhood collinearity scores, For temperature parameters; The mapping matrix after quality control The metadata of homologous gene clusters is encapsulated into a structured mapping library. ;in A set of gene function annotation vectors for the target species. This is the set of pre-trained representation vectors of the source species' genes. To divide homologous gene clusters into sets, It is a collection of homologous gene clusters; In actual evaluation, the gene neighborhood collinearity score is calculated by the MCScanX algorithm based on the conservative arrangement of gene pairs in the chromosomal location neighborhood; the temperature parameter controls the distribution sharpness; the expansion parameter is 2, which is used to control the intra-cluster cohesion and inter-cluster separation. Middle elements This represents the homology confidence between target gene j and source gene i; non-homologous pairs are assigned a value of 0. Arabidopsis thaliana and rice (model crops with well-studied salt tolerance and high gene homology) were selected as source species, and whole-genome protein coding sequence and functional annotation data were collected. Sequence preprocessing: Low-complexity region filtering → signal peptide cleavage to obtain a standardized genome dataset; Smith-Waltman algorithm with a gap penalty of 2 and an amino acid substitution score matrix of BLOSUM62; candidate homologous pairs with similarity > 0.75 are retained; Markov clustering stops when the entropy change threshold is < 0.01, resulting in 128 homologous gene clusters; orthologous homology confidence threshold of 0.6; chromosome collinearity constraints are calculated by the MCScanX algorithm, finally generating a structured mapping library; By transferring knowledge of salt-tolerant genes from Arabidopsis thaliana and rice to rapeseed through a mapping library, the model's feature expression capability was improved by 32% in small sample scenarios.

[0019] In this embodiment, the method for constructing a policy optimization mechanism that coordinates outer-layer meta-learning and inner-layer reinforcement learning includes: Outer Meta-Learning: Employing gradient-based efficient meta-learning on a task distribution composed of multiple salt tolerance trait-related tasks. Learning model initial parameters ,in To execute on the k-th task Task-specific parameters after gradient descent. The meta-learning rate, The total number of tasks. This represents the number of gradient descent steps in the inner loop. These are the initial parameters; Inner reinforcement learning: Each training batch is treated as a Markov decision process, and the training strategy is dynamically adjusted using a proximal policy optimization algorithm; where the state space... The joint characterization of the distribution entropy of the current batch of samples Conflict angles of gradients for each mode of loss Validation set performance degradation rate and sample class imbalance The components are assembled, and the dimensions of the assembly are unified. Action space Including the sampling probability adjustment vector for each modality Loss weighting coefficient and learning rate scaling factor Policy Network: A two-layer MLP is used to approximate the policy function. and value function Reward function Improved accuracy on the comprehensive validation set The sample efficiency gain and training stability are expressed as follows: ;in For classification loss weights, For biological semantic contrast loss weights, For L2 regularization loss weights, For the policy function, Let v be the reward value for the v-th training step. This represents the number of labeled samples in the current batch. These are the weighting coefficients for the sample efficiency term. To train the weight coefficients of the stability term, The variance of the training loss; At each training step, the policy network outputs an action. Dynamic adjustment: based on the status The sampling probabilities of each modality are reweighted, prioritizing the sampling of difficult samples with high information gain, by adjusting... Balance the classification loss, biological semantic contrast loss, and L2 regularization loss, and dynamically scale the learning rate according to the training phase; The policy network is trained by optimizing the pruning objective function using a proximal strategy. The expression is: ; in For probability ratios, For generalized advantage estimation, The objective function for optimizing the near-end strategy is... For the expected value of all training steps, To truncate hyperparameters, This is a truncation function; In the actual evaluation, five sub-tasks were divided according to the salt concentration gradient: 0.3%, 0.4%, 0.5%, 0.6%, and 0.7% salt content; the meta-learning rate was 0.01, the number of inner loop gradient descent steps was 5, and the total number of tasks was 5; the goal was to learn general initial parameters to adapt to different salt and alkali stress scenarios. The concatenation dimension of the state space is 64. It is 0.3. The hyperparameter is set to 0.2, prioritizing strategies with high accuracy, high sample efficiency, and low training fluctuation; the truncation hyperparameter is 0.2; the model automatically adjusts the training strategy, increasing the sampling probability of difficult samples by 40%, reducing the training loss variance by 28%, and improving the convergence speed by 25%.

[0020] In this embodiment, the method for constructing the rapeseed salt-alkali tolerance prediction model includes: Using 7-tuple A predictive model for salt and alkali tolerance traits in rapeseed was developed, and differential parameter initialization was performed on each component of the seven-tuple; among which... It is a stereo encoder. It is a genotype encoder. For environmental encoders, For cross-modal attention fusion module, For cross-species migration module, For policy networks, For prediction head; The rapeseed salt-alkali tolerance prediction model includes encoder parameter initialization and prior injection, cross-species migration mechanism embedded encoder, and meta-reinforcement collaborative training algorithm; Encoder parameter initialization and prior injection: semantic anchor matrix of salt tolerance gene pathway The embedded genotype encoder optimizes the objective by updating the semantic anchor matrix of salt tolerance gene pathways via exponential moving average during training. ,in For the first The semantic anchor vector of the p-th path during training, The momentum coefficient, For the first The semantic anchor vector of the p-th path during training, For the first The subset of samples annotated to the p-th path during training. For the first The genotype feature vector of each gene sample output by the model in the current round of training; Cross-species migration mechanisms embedded in encoders: in the genotype coding branch After the graph attention layer outputs, a cross-species transfer module is embedded. : Hidden features of the target gene are mapped using a library Retrieval of source species pre-trained representations Calculate the gate vector and output the transfer enhancement feature. ;in For the target species after interspecies migration The hidden layer feature vectors of each gene in the encoder output. For the target species The first gene and its source species The reliability of homology relationships among individual genes A learnable transfer matrix, For the source species Pre-trained representation of each gene, This represents the number of homologous genes retrieved in the source species. It is the sigmoid activation function. The learnable gated weight matrix, For adaptive gating vectors, For element-wise multiplication, For the target species The hidden layer feature vectors of each gene in the encoder output; Meta-reinforcement collaborative training algorithm: Defining the task distribution for predicting rapeseed salt and alkali tolerance traits Each task corresponds to a specific salinization stress condition, and each task includes a support set and a query set. The outer loop objective of the meta-learning is to find the optimal initial parameters, and a first-order approximation model-independent meta-learning is adopted. ;in These are the initial parameters. For the model in the task The parameters on the support set after n steps of gradient descent The total number of tasks. The meta-learning rate; Each training batch is treated as a time step in a Markov decision process. An inner loop for policy optimization is constructed, and a state vector is built at each training step. An action-space policy network is defined. Output Action ;in Let v be the state vector at the v-th training step. For the action of the v-th training step, Let v be the learning rate scaling factor for the v-th training step. Assign coefficients to the loss weights for the v-th training step. This is the sampling probability adjustment vector for the v-th training step; Given a reward function The network parameters of the update strategy are optimized using a near-end strategy, based on task distribution. Randomly sample K tasks, and for each task Execute inner layer adaptation, during policy network Each step outputs an action. Adjust training hyperparameters and evaluate adapted parameters on the query set. Calculate task loss and reward, and aggregate all task gradients to update the initial parameters. Network parameters based on cumulative trajectory update strategy ; Predicting Head A three-layer fully connected network is adopted, with the hidden layer dimensions being as follows: BatchNorm and LeakyReLU activations are inserted between layers, and the output layer uses Sigmoid activation to constrain the predicted values ​​to [0,1]. The expression is as follows: ;in Here are the learnable parameters of the prediction head; where The dimension of the joint representation of z after cross-modal fusion; The composite loss function is composed of three sub-losses dynamically weighted by the policy network. ,in For classifying losses, For biological semantic contrast loss, For parameter regularization loss, For classification weights, For biological semantic contrast weights, Regularize the weights for the parameters; Training is stopped and the target model is output when the mean absolute error of the validation set does not decrease for 10 consecutive meta-iterations. In the actual evaluation, the phenotypic encoder initialization was performed using ImageNet large-scale visual pre-trained weights. The MobileViT backbone network retains the ability to extract visual features from the lower layers; the top-level classification head is randomly initialized. Genotype encoder initialization: GAT layer parameters were initialized using a Xavier uniform distribution; mapping library The homology confidence matrix in As a topological constraint of the graph structure, it is directly fixed and does not participate in gradient backpropagation; Environmental encoder initialization: The TCN causal convolution kernel parameters are initialized using a He normal distribution, and the dilation factor is calculated using an exponential sequence. Preset, For network layer indexing; A three-layer fully connected network [384,192,1] is used with BatchNorm+LeakyReLU activation. The output layer Sigmoid is normalized to [0,1]. The weights of the composite loss are dynamically adjusted by the policy network. The Monte Carlo dropout method was used, with 20 forward propagations and a model accuracy prior of 10. The dynamic variance threshold was 0.9, the q quantile was 0.9, the early relaxation coefficient was 0.5, the decay rate was 0.1, and the rare class correction intensity was 0.3. In this embodiment, 12 difficult samples with prediction variance exceeding the threshold were included in the training set, further improving the model prediction accuracy by 8.5%.

[0021] In this embodiment, the method for obtaining the prediction variance based on the prediction data includes: Multi-source heterogeneous data of the rapeseed variety to be predicted are input into the trained three-modal coupled encoder to obtain joint representations. Input prediction header: ;in The comprehensive index of salt and alkali tolerance. It is a multilayer perceptron; Forward propagation using the Monte Carlo dropout method Next, calculate the prediction variance. : ; ; in To predict the mean, As a priori for model accuracy, Forward propagation count, is the comprehensive index of salt and alkali tolerance during the t-th forward propagation.

[0022] In this embodiment, the expression for the variance threshold is: ; in Let be the variance threshold at the t-th forward propagation. This is the global conservative coefficient. To validate the q-quantile function of the variance distribution predicted by the validation set, The set of predicted variances for the validation set samples. This represents the relaxation factor during early training. The dynamic decay rate is t, where t is the current training round. To adjust the strength for category scarcity, It is the numerical stability constant. For varieties Category The current number of labeled samples; In actual evaluation, the coefficient of determination for predicting salt-alkali tolerance traits in 120 rapeseed varieties was 0.91, and the mean absolute error (MAE) was 0.042, which is much higher than that of the traditional model (coefficient of determination is 0.72). The model was trained with only 30 labeled samples, and the generalization accuracy reached 0.86, solving the problem of scarce labeled data for rapeseed salt-alkali tolerance. Eleven highly salt-alkali tolerant rapeseed varieties were quickly screened, with a 92.3% agreement rate with field identification results, shortening the breeding cycle by more than 60%. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a deep learning-based prediction model for salt-alkali tolerance traits in rapeseed, characterized in that, Includes the following steps: Multimodal data of a predefined rapeseed region is collected and preprocessed; the multimodal data includes phenotypic modality data, genotypic modality data, and environmental modality data. Based on the multimodal data, a three-modal coupled encoder with biological prior constraints is constructed. For the small sample scenario of rapeseed, a cross-species migration mechanism based on gene homology mapping is constructed to obtain a mapping library. A policy optimization mechanism that combines outer-layer meta-learning and inner-layer reinforcement learning is constructed, and a rapeseed salt-alkali tolerance trait prediction model is built based on the three-modal coupled encoder, the mapping library, and the policy optimization mechanism. The data to be predicted is input into the rapeseed salt-alkali tolerance trait prediction model to obtain prediction data. The prediction variance is obtained based on the prediction data. When the prediction variance is greater than the variance threshold, the active learning sampling mechanism is triggered to include the current sample into the training support set and output the target model.

2. The method for constructing a deep learning-based prediction model for rapeseed salt-alkali tolerance according to claim 1, characterized in that, A method for constructing a biologically prior-constrained trimodal coupled encoder based on the multimodal data includes: Obtain the phenotypic tensor, genotype matrix, and environment tensor, and construct a three-modal coupled encoder using phenotypic coding branch, genotype coding branch, and environment coding branch; Phenotypic encoding branch: Employs a lightweight multispectral vision Transformer, inputting phenotypic tensors. By employing patch embedding, reverse residual bottleneck, and global attention mechanisms, phenotypic latent vectors are output. Genotype coding branch: A graph attention network is constructed based on the gene interaction network, where nodes are single nucleotide polymorphism sites and edges represent linkage disequilibrium. The input is the genotype matrix. Output genotype latent vector Environment encoding branch: Employs a temporal convolutional network to encode the dynamics of the environment, with the input environment tensor... Output environment latent vector ; Constructing a semantic anchor matrix for salt and alkali tolerance gene pathways ,in For the prior number of paths, As a dimension, each row Let p be the semantic center vector of the p-th pathway; define the biological semantic contrast loss. : ； in For varieties The set of salt-alkali tolerance pathways annotated, For a set of irrelevant paths, For the interval hyperparameter, For variety index, For varieties Genotype latent vector, For set The number of intermediate pathways, The square of the L2 norm of the vector; The phenotypic latent vector, genotype latent vector, and environmental latent vector are fused into a joint representation through a cross-modal attention fusion layer, expressed as: ； ； in For queryable vectors, For modal adaptive weights, For joint representation vector, For modal indexing, For modal adaptive weights, Let be the learnable weight matrix for the m-th mode. This is the transpose of the queryable vector. for.

3. The method for constructing a deep learning-based prediction model for rapeseed salt-alkali tolerance according to claim 1, characterized in that, Methods for constructing cross-species migration mechanism mapping libraries based on gene homology mapping include: We collected whole-genome protein-coding sequences and functional annotation data of source and target species, and performed low-complexity region filtering and signal peptide splicing preprocessing on the protein sequence sets of source and target species to obtain a standardized genome dataset. A dynamic programming algorithm is used to perform full-pair whole-sequence alignment between the source and target species to construct an initial similarity score matrix. The Smith-Waltman algorithm is then used to calculate the global alignment score between any source and target sequences. ; For each pair Perform bidirectional optimal comparison and screening: forward reverse Candidate homologous pairs are retained when both bidirectional optimal conditions are met. The scores of candidate homologous pairs are normalized. Construct a weighted homology graph Vertex set edge set Edge weight matrix If an edge exists, then the element Otherwise, it is 0; Markov clustering algorithm is used to analyze the weighted homology graph. Homologous gene clusters were divided, and a transition probability matrix was constructed. Execute extension With expansion The process stops when the entropy change of the transition probability matrix is ​​less than the entropy change threshold, thus obtaining a set of homologous gene clusters. Calculate homologous gene clusters Intraspecific target species genes Genes of endogenous species in the collection of homologous gene clusters direct homology confidence between The orthologous homology confidence scores within each cluster are aggregated into a global gene homology mapping matrix. ; Only source gene mappings with a confidence level greater than a confidence threshold are retained, and Softmax normalization of the global gene homology mapping matrix is ​​performed along the source gene dimension. For gene clusters with many-to-many mappings, a chromosomal neighborhood collinearity constraint is introduced to obtain the mapping matrix. ; The mapping matrix after quality control The metadata of homologous gene clusters is encapsulated into a structured mapping library. .

4. The method for constructing a deep learning-based prediction model for rapeseed salt-alkali tolerance according to claim 1, characterized in that, Methods for constructing a policy optimization mechanism that coordinates outer-layer meta-learning and inner-layer reinforcement learning include: Outer Meta-Learning: Employing gradient-based efficient meta-learning on a task distribution composed of multiple salt tolerance trait-related tasks. Learning model initial parameters ; Inner reinforcement learning: Each training batch is treated as a Markov decision process, and the training strategy is dynamically adjusted using a proximal policy optimization algorithm; where the state space... The joint characterization of the distribution entropy of the current batch of samples Conflict angles of gradients for each mode of loss Validation set performance degradation rate and sample class imbalance Composed of splicing elements, with unified splicing dimensions. Action space Including the sampling probability adjustment vector for each modality Loss weighting coefficient and learning rate scaling factor Policy Network: A two-layer MLP is used to approximate the policy function. and value function Reward function Improved accuracy on the comprehensive validation set The sample efficiency gain and training stability are expressed as follows: ; At each training step, the policy network outputs an action. Dynamic adjustment: based on the status The sampling probabilities of each modality are reweighted, prioritizing the sampling of difficult samples with high information gain, by adjusting... Balance the classification loss, biological semantic contrast loss, and L2 regularization loss, and dynamically scale the learning rate according to the training phase; The policy network is trained by optimizing the pruning objective function using a proximal strategy. The expression is: ； in For probability ratios, For generalized advantage estimation, The objective function for optimizing the near-end strategy is... For the expected value of all training steps, To truncate hyperparameters, This is a truncation function.

5. The method for constructing a deep learning-based prediction model for rapeseed salt-alkali tolerance according to claim 1, characterized in that, The method for constructing the rapeseed salt-alkali tolerance prediction model includes: Using 7-tuple Characterize the salt and alkali tolerance prediction model of rapeseed by performing differential parameter initialization on each component of the seven-tuple; The rapeseed salt-alkali tolerance prediction model includes encoder parameter initialization and prior injection, cross-species migration mechanism embedded encoder, and meta-reinforcement collaborative training algorithm; Encoder parameter initialization and prior injection: semantic anchor matrix of salt tolerance gene pathway The embedded genotype encoder optimizes the objective by updating the semantic anchor matrix of salt tolerance gene pathways via exponential moving average during training. ; Cross-species migration mechanisms embedded in encoders: in the genotype coding branch After the graph attention layer outputs, a cross-species transfer module is embedded. : Hidden features of the target gene are mapped using a library Retrieval of source species pre-trained representations Calculate the gate vector and output the transfer enhancement feature. ; Meta-reinforcement collaborative training algorithm: Defining the task distribution for predicting rapeseed salt and alkali tolerance traits Each task corresponds to a specific salinization stress condition, and each task includes a support set and a query set. The outer loop objective of the meta-learning is to find the optimal initial parameters, and a first-order approximation model-independent meta-learning is adopted. ; Each training batch is treated as a time step in a Markov decision process. An inner loop for policy optimization is constructed, and a state vector is built at each training step. An action-space policy network is defined. Output Action ; Given a reward function The network parameters of the update strategy are optimized using a near-end strategy, based on task distribution. Randomly sample K tasks, and for each task Execute inner layer adaptation, during policy network Each step outputs an action. Adjust training hyperparameters and evaluate adapted parameters on the query set. Calculate task loss and reward, and aggregate all task gradients to update the initial parameters. Network parameters based on cumulative trajectory update strategy ; Predicting Head A three-layer fully connected network is adopted, with the hidden layer dimensions being as follows: BatchNorm and LeakyReLU activations are inserted between layers, and the output layer uses Sigmoid activation to constrain the predicted values ​​to [0,1]. The expression is as follows: ; The composite loss function is composed of three sub-losses dynamically weighted by the policy network. ; Training is stopped and the target model is output when the mean absolute error of the validation set does not decrease for 10 consecutive meta-iterations.

6. The method for constructing a deep learning-based prediction model for rapeseed salt-alkali tolerance according to claim 1, characterized in that, A method for obtaining the prediction variance based on the prediction data includes: Multi-source heterogeneous data of the rapeseed variety to be predicted are input into the trained three-modal coupled encoder to obtain joint representations. Input prediction header: ;in The comprehensive index of salt and alkali tolerance. It is a multilayer perceptron; Forward propagation using the Monte Carlo dropout method Next, calculate the prediction variance. : ； ； in To predict the mean, As a priori for model accuracy, Forward propagation count, is the comprehensive index of salt and alkali tolerance during the t-th forward propagation.

7. The method for constructing a deep learning-based prediction model for rapeseed salt-alkali tolerance according to claim 1, characterized in that, The expression for the variance threshold is: ； in Let be the variance threshold at the t-th forward propagation. This is the global conservative coefficient. To validate the q-quantile function of the variance distribution predicted by the validation set, The set of predicted variances for the validation set samples. This represents the relaxation factor during early training. The dynamic decay rate is t, where t is the current training round. To adjust the strength for category scarcity, It is the numerical stability constant. For varieties Category The current number of labeled samples.

Images