Aquatic genetic trait prediction method and system based on multi semantic feature fusion and knowledge distillation

Abstract

This application provides a method and system for predicting aquatic genetic traits based on multi-semantic feature fusion and knowledge distillation, relating to the field of biogenetics. The method includes: acquiring SNP data and phenotypic labels; constructing six-dimensional semantic features and using CLIP cross-modal contrastive learning to obtain enhanced feature embeddings; constructing a GIA-Net teacher network, whose graph branch captures local linkage features and sequence branch captures global long-range dependency features based on Mamba, dynamically fusing these features to generate prediction results and site importance scores; constructing a lightweight BioMamba student network, trained through multi-level knowledge distillation; and using the trained student network to predict traits. This application achieves a systematic integration of computational efficiency and genetic analysis capabilities, providing an efficient and secure digital decision-making tool for disease-resistant breeding of plants and animals and for mining key variation sites.

Description

Technical Field

[0001] This application relates to the field of biogenetic technology, and in particular to a method and system for predicting aquatic genetic traits based on multi-semantic feature fusion and knowledge distillation. Background Technology

[0002] Against the backdrop of global climate change and population growth, building a sustainable food supply system has become a major challenge. Aquaculture, as the fastest-growing food production sector, is key to providing high-quality animal protein. Currently, global aquaculture production has historically surpassed that of wild-caught fisheries, becoming a core engine driving the blue transition. However, the rapid expansion of the industry has led to increased ecological pressure on the aquaculture environment and the proliferation of pathogenic microorganisms. In recent years, emerging and recurring diseases have frequently broken out worldwide, and drug abuse has triggered biosecurity crises such as environmental residues and drug resistance, hindering the industry's green and sustainable development.

[0003] To address the challenge of disease, breeding superior varieties with enduring disease resistance is widely recognized as the most fundamental solution. Traditional breeding relies on phenotypic and pedigree information, but disease resistance traits are characterized by low heritability and difficulty in measurement, leading to slow genetic progress. With the development of sequencing technology, Genomic Selection (GS) technology has emerged. This technology achieves early selection through molecular markers covering the entire genome, significantly shortening generation intervals and improving breeding accuracy. Its core step is to accurately identify "disease resistance" or "susceptibility" information based on individual genotypic differences, providing crucial support for elucidating the genetic basis of complex diseases in aquatic animals.

[0004] While GS technology has made breakthroughs in some economic traits, its core algorithms still heavily rely on GBLUP or Bayesian statistical methods. These classic models typically assume that traits are the sum of linear additive effects of a large number of minor genes, ignoring the widespread nonlinear regulation in organisms. Modern systems biology research shows that complex traits such as disease resistance are profoundly influenced by nonlinear interactions between genes (i.e., epistatic effects). Traditional linear models struggle to effectively capture long-distance, nonlinear interaction characteristics in the genome, leading to the loss of some key genetic variation information and limiting further breakthroughs in prediction accuracy.

[0005] In recent years, deep learning technology has shown significant advantages in mining potential high-order gene interactions due to its powerful feature extraction capabilities. For example, convolutional neural networks (CNNs) study nonlinear mappings through local features, while long-sequence modeling networks such as Caduceus attempt to capture long-range dependency information. However, when dealing with ultra-long whole-genome sequences and actual breeding scenarios, existing research still faces three deep-seated technical bottlenecks: (1) The problem of semantic missing site information. Existing models mostly treat SNPs as isolated numerical indices, stripping away the underlying biological background such as functional annotations, linkage disequilibrium (LD), and GWAS priors. This semantically deficient feature representation makes it difficult for models to extract association features that conform to genetic logic from sparse genomic signals. (2) Bottleneck in gene interaction mining. Disease resistance traits are driven by both local LD ​​and global long-range dependencies. Existing simple convolutional or attention architectures are difficult to capture both local associations and global nonlinear interactions at a single scale, limiting the depth of analysis of complex genetic structures. (3) The contradiction between model complexity and practical deployment. The high dimensionality of whole-genome data leads to redundant parameters and high inference latency in deep learning architectures. Faced with the need for high-throughput real-time estimation of tens of thousands of individuals, overly complex models are difficult to scale up in resource-constrained environments. Summary of the Invention

[0006] This application provides a method and system for predicting aquatic genetic traits based on multi-semantic feature fusion and knowledge distillation. It aims to solve problems such as insufficient feature representation ability due to the lack of biological semantics of locus information, the difficulty of traditional models in simultaneously capturing local linkage and global long-range nonlinear interactions, and the huge computational cost of deep learning models under high-dimensional data, which makes them difficult to deploy on a large scale. By mapping discrete SNP loci to a continuous semantic feature space that integrates multi-source biological priors through cross-modal contrastive learning, a graph-sequence dual-branch teacher network is constructed to deeply mine epistatic interactions, and multi-level knowledge distillation is used to transfer complex discrimination logic to a lightweight student network with linear complexity. This achieves a systematic and synergistic improvement in prediction accuracy, computational efficiency and genetic interpretability, providing an efficient and safe digital decision-making tool for disease-resistant aquatic breeding.

[0007] In a first aspect, this application provides a method for predicting aquatic genetic traits based on multi-semantic feature fusion and knowledge distillation, the method comprising: Obtain whole-genome SNP data and corresponding disease resistance / susceptibility phenotype tags for aquatic animals; We construct multi-semantic features and utilize a contrastive language-image pre-training architecture to map the original SNP data to a continuous semantic feature space that incorporates biological priors through cross-modal contrastive learning, thereby obtaining enhanced genomic feature embeddings. A teacher network is constructed, which adopts a dual-branch architecture, including a graph branch and a sequence branch. The enhanced genomic features are embedded and input into the graph branch to obtain local structural features and the sequence branch to obtain a global sequence feature matrix. The local structural features and the global sequence feature matrix are then fused through a dynamic fusion module to generate teacher network prediction results and site importance scores. A lightweight student network is constructed; a multi-level knowledge distillation strategy is adopted, and the student network is trained using the prediction results generated by the teacher network and the site importance scores to obtain the trained lightweight prediction model BioMamba. The trained lightweight prediction model BioMamba is used to predict traits from the genomic SNP data to be predicted, and outputs the classification results of disease resistance or susceptibility.

[0008] Furthermore, the multi-semantic features include six dimensions of genomic semantic features, namely, genotype construction semantic features, statistical prior semantic features, regulatory hotspot semantic features, gene density semantic features, haplotype structure semantic features, and quality mask semantic features. Among them, genotype semantic features The calculation formula is as follows: (1) in, For the first i The original genotype values ​​of each sample; Statistical prior semantic features The calculation method is as follows: Determine the sample n phenotypic probability Let be a nonlinear function with respect to the parameter to be determined: (2) in, For the first i The conditional probability that a sample will exhibit a specific phenotype (such as disease resistance) given a given genotype. For the first n The true phenotypic label of each sample (e.g., 1 represents disease resistance, 0 represents susceptibility), This represents the SNP effect value. It is a natural exponential function. This is the intercept term of the logistic regression model. For the first n The genotype of each sample For the first c One principal component covariate, For the first c The effect coefficients of the principal component covariates, The total number of principal component covariates; Construct the Firth penalized log-likelihood function : (3) in, This is the determinant of the Fisher information matrix; The unbiased estimator that makes the function reach its global maximum is obtained by solving the Newton-Raphson iterative algorithm. : (4) in, Firth penalized log-likelihood function, It is a vector consisting of all parameters to be estimated in the model (including intercept term, principal component covariate coefficients, and SNP effect values); Normalized SNP effect values As prior weights The statistical prior semantic features are calculated using the following formula, corresponding to the statistical prior channels in the feature matrix. : (5) Regulation of hotspot semantic features The calculation formula is as follows: (6) in, The 90th percentile threshold; This represents the frequency of the second allele. To examine allele frequencies; min is a minimum value function; It is a mean function; Genotype data of the training set after imputing missing values; Semantic features of gene density The calculation formula is as follows: (7) in, For the first i The total number of variations within the neighborhood of each SNP For the first k The physical location of each SNP For the first i The physical location of each SNP This represents the minimum total number of variants across all SNP neighborhoods. The maximum value of the total number of variants in the neighborhood of all SNPs. To prevent extremely small constants with a denominator of zero, The total number of SNPs on a chromosome or sequence; Haplotype structural semantic features The calculation formula is as follows: (8) in, For LD blocks, For the SNP index within the block, For the currently evaluated SNP index; Quality mask semantic features The calculation formula is as follows: (9) in, This encodes the original genotype at the i-th locus. The indicator function represents the semantic features of the quality mask when the genotype data is valid (i.e., a normal genotype with a value of 0, 1, or 2). The value is 1.

[0009] Furthermore, methods for contrastive learning using a contrastive language-image pre-training architecture include: The six-dimensional genomic semantic features are reconstructed into an RGB three-channel image using the following formula: (10) in, The pixel value for the red channel. The pixel value is for the green channel. The pixel value is for the blue channel. Genotype semantic features To statistically analyze prior semantic features, To regulate the semantic features of hotspots, For gene density semantic features, It is a semantic feature of haplotype structure. For quality mask semantic features; By optimizing the symmetric InfoNCE loss function Align the two modalities in a shared feature space, where Represented as: (11) in, The first extracted by the visual encoder i One genome modality feature, The first one extracted by the text encoder i A semantic modality feature, For temperature coefficient, This refers to the batch size.

[0010] Furthermore, the methods for embedding the enhanced genomic features into the input graph branch to obtain local structural features include: A strategy combining sliding window and linkage disequilibrium is used to construct a genome heterogeneity map, wherein the first... k Nodes SNP index set Represented as: (12) in, For SNP site indexing, The sliding step size, For physical coordinates, This refers to the size of the sliding window; Edge weights between nodes Calculated using the following formula: (13) Where max is the maximum value function. For nodes and The LD weights between them; The spatial proximity weight is calculated using the following formula: (14) in, and The center physical coordinates of the two nodes; Initial feature vector of a node Calculated using the following formula: (15) in, For the embedding of the i-th enhanced genomic feature, For nodes The set of SNP indexes included; Based on the aforementioned genomic heterogeneity map, local genetic patterns are mined using graph isomorphism networks and dynamic graph attention networks, and a Mamba module is introduced to reconstruct linear order dependencies. The local structural features are then output through a hybrid expert module.

[0011] Furthermore, based on the aforementioned genomic heterogeneity map, local genetic patterns are mined using graph isomorphism networks and dynamic graph attention networks, and a Mamba module is introduced to reconstruct linear order dependencies. The local structural features are then output through a hybrid expert module, specifically including... In the graph isomorphic network, the first layer hidden features The update formula is: (16) in, For learnable scalar parameters; For layer normalization function; It is a multilayer perceptron; For nodes i The initial feature vector; For nodes i The set of neighboring nodes; For neighboring nodes j The initial feature vector; In the dynamic graph attention network, nodes i with neighbors j Dynamic attention coefficient The calculation formula is: (17) in, The weight matrix is ​​a learnable matrix. For nodes j The first layer of latent features This is a matrix transpose operation. This is a learnable attention weight vector; Local explicit structural features after Softmax normalization The update formula is as follows: (18) in, The dynamic attention coefficients in formula (17) Weights obtained after Softmax normalization; After importing the Mamba module, the hidden state is... and the corrected node features The update formula is as follows: (19) in, , For time-varying parameters, To output the projection matrix, This is the hidden state from the previous moment. For the sorted numbert The input feature vector of each node; The hybrid expert module has features for any node. A Top-K gating mechanism incorporating standard normal Gaussian noise is introduced to calculate the matching confidence with each expert network: (20) in, The weight distribution of the gating network output. To retain the top K maximum values, For the learnable weight matrix of the gated network, The noise is a standard normally distributed Gaussian noise. Local structural features are calculated using the following formula. : (twenty one) in, For the first e The weighting of each expert. For expert network indexing, The total number of expert networks, For the first e The output of an expert network.

[0012] Furthermore, the method of embedding the enhanced genomic features into the sequence branch to obtain the global sequence feature matrix includes: The enhanced genomic features are embedded and processed into a one-dimensional continuous sequence input sequence branch; The sequence branches adopt the Mamba architecture, including parallel feature main branches and gated branches; In the main feature branch, the input sequence extracts the local physical associations between adjacent sites through one-dimensional convolution, and after being processed by the SiLU activation function, it enters the state space model operator. In the state-space model operator, time-varying parameters are dynamically generated through a selective scan mechanism, wherein the discretized state transition matrix... With projection matrix The calculation formula is: (twenty two) in, The step size is used for discretizing continuous systems; its function is to transform continuous-time parameters into discrete state transition matrices. and projection matrix , It is the identity matrix. For time step, It is a continuous state transition matrix. It is a continuous projection matrix; After discretization, the hidden state The recursive evolution and output calculation formulas are shown below: (twenty three) in, This represents the discrete hidden state at the previous time step. The input sequence features at the current time step; In the gated branch, the input features are linearly projected to generate a control signal; The control signal is multiplied element-wise with the output of the main branch SSM operator to obtain the gated modulation features. The gated and modulated features are processed sequentially through an output projection layer and a residual connection. The residual connection adds the output of the output projection layer to the original input of the sequence branch to obtain the global sequence feature matrix.

[0013] Furthermore, the local structural features and global sequence feature matrix are fused through a dynamic fusion module to generate teacher network prediction results and site importance scores, specifically including: Map local structural features to query vectors The global sequence feature matrix is ​​mapped to the key. Sum The output of each attention head is calculated through a multi-head cross-attention mechanism; wherein, the calculation formula of the multi-head cross-attention mechanism is: (twenty four) in, For the output of each attention head, For attention weights, T This is a matrix transpose operation. The feature dimension of the key vector; The outputs of each attention head are linearly projected through a feature reconstruction layer (FR), and then superimposed with residual connections and layer normalization to obtain an enhanced feature matrix containing local variation signals and macroscopic environmental information. ; Enhance the feature matrix The inputs are respectively fed into a first task branch and a second task branch; wherein, the first task branch is used to process the enhanced feature matrix. The downstream classifier performs disease resistance trait prediction and obtains the predicted disease resistance trait as the teacher network prediction result; the second task branch is used to perform the prediction based on the enhanced feature matrix. The importance score of a site is calculated using the following formula: (25) in, For the first i Locus importance score of each SNP , For a learnable projection matrix, For activation function, For macro-block weights, For Hadama accumulation, For the first i Enhanced feature matrix of SNPs.

[0014] Furthermore, after obtaining the site importance score, the method also includes constructing a genome-wide SNP-SNP interaction matrix based on the site importance score. M : (26) in, This represents the importance score of the locus.

[0015] Furthermore, in the multi-level knowledge distillation strategy, the total loss function used is... Represented as: (27) in, , To balance the hyperparameters of the component weights, For the loss of mission supervision, For response stage distillation loss, Characteristic stage distillation loss; The task supervision loss Calculated using the following formula: (28) in, B For batch size, This is the true phenotypic label for the sample. To predict the probability distribution; The response-level distillation loss Calculated using the following formula: (29) in, Temperature coefficient; Output scores to students online; Scoring of the output of the teacher network; The feature-level attention distillation loss Calculated using the following formula: (30) in, For sequence length, and The first generation generated by the teacher network and the student network respectively i Importance score of each site.

[0016] Secondly, this application provides an aquatic genetic trait prediction system based on multi-semantic feature fusion and knowledge distillation, used to implement the aquatic genetic trait prediction method based on multi-semantic feature fusion and knowledge distillation as described above, the system comprising: The data acquisition module is configured to acquire whole-genome SNP data of aquatic animals and corresponding disease resistance / susceptibility phenotype tags; The data augmentation module is configured to construct multi-semantic features and utilize a contrastive language-image pre-training architecture to map the original SNP data to a continuous semantic feature space that incorporates biological priors through cross-modal contrastive learning, thereby obtaining enhanced genomic feature embeddings. The teacher network construction module is configured to construct a teacher network, which adopts a dual-branch architecture, including a graph branch and a sequence branch. The enhanced genomic features are embedded and input into the graph branch to obtain local structural features, and into the sequence branch to obtain a global sequence feature matrix. The local structural features and the global sequence feature matrix are then fused through a dynamic fusion module to generate teacher network prediction results and site importance scores. The student network training module is configured to construct a lightweight student network; a multi-level knowledge distillation strategy is adopted to train the student network using the prediction results and site importance scores generated by the teacher network, resulting in the trained lightweight prediction model BioMamba. The classification prediction module is configured to use the trained lightweight prediction model BioMamba to predict traits in the genomic SNP data to be predicted, and output the classification result of disease resistance or susceptibility.

[0017] The aquatic genetic trait prediction method and system based on multi-semantic feature fusion and knowledge distillation provided in this application have at least the following beneficial effects: To address the bottlenecks in genome-wide prediction, such as missing semantic features, high-dimensional feature redundancy, difficulties in modeling long sequences, and lack of model interpretability, this application constructs a six-dimensional semantic feature set including SNPV, GWASV, REV, GDV, HSV, and DQMV using the Geno-CLIP module. It then utilizes cross-modal contrastive learning to align the genome sequence with higher-order biological semantics, filtering out genetic background noise and reshaping a high signal-to-noise ratio feature base in the early stages of feature extraction. The teacher network GIA-Net adopts a graph-sequence dual-branch architecture, with the graph branch capturing sequences through heterogeneous graphs, GIN, GATv2, Mamba, and MoE. This paper obtains local linkage features, uses sequence branching based on linear complexity SSM to capture global long-range dependency features, and employs a dynamic fusion module to mine epistatic effects using multi-head cross-attention. Based on MDIS scores, it accurately identifies core candidate loci on key chromosomes and the nonlinear interaction network between loci. Through multi-level knowledge distillation, the response-level and feature-level discrimination logic of the teacher network is transferred to the lightweight student network BioMamba, significantly reducing the number of parameters and inference overhead while outperforming existing benchmark models in all core evaluation metrics. Ablation experiments verify the necessity of the collaborative design of pre-training, graph space architecture, and knowledge distillation. The MC Dropout mechanism further quantifies prediction uncertainty to identify high-risk samples, and internal attention combined with SHAP analysis accurately locates SNP loci driving phenotypic variations and restores allele additive effects. Therefore, this application achieves a balance between lightweight design, high interpretability, and computational security while ensuring trait prediction accuracy, providing an efficient and reliable digital decision-making tool for disease-resistant breeding in aquaculture and plants and animals. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] Figure 1 A BioMamba architecture diagram for implementing a method for predicting aquatic genetic traits based on multi-semantic feature fusion and knowledge distillation, provided in this application embodiment; Figure 2 A flowchart illustrating a method for predicting aquatic genetic traits based on multi-semantic feature fusion and knowledge distillation, provided in this application embodiment; Figure 3The heatmaps of comparative experimental results provided for embodiments of this application include: (a) comparison results of 8 methods on the ACC metric for 12 datasets; (b) comparison results of 8 methods on the AUC metric for 12 datasets; (c) comparison results of 8 methods on the Presion metric for 12 datasets; (d) comparison results of 8 methods on the Recall metric for 12 datasets; (e) comparison results of 8 methods on the F1 score for 12 datasets; and (f) comparison results of 8 methods on the MCC metric for 12 datasets. Figure 4 The Geno-CLIP semantic alignment distribution map provided for the embodiments of this application on four representative datasets verifies the model's ability to map cross-species genetic signals and biological semantics in a high-dimensional consistency. Figure 5 The image shows a comparison of sample distribution and separation before and after Geno-CLIP feature enhancement, provided for embodiments of this application; wherein (a) is the original SNP genotype feature distribution; and (b) is the SNP genotype feature distribution after Geno-CLIP training. Figure 6 The evolution trajectory of the BioMamba feature space with training rounds provided in this application embodiment; wherein (a) after the first round of training; (b) after the second round of training; (c) after the third round of training; (d) after the fourth round of training; Figure 7 The feature space evolution trajectory diagram provided in this application embodiment records the feature evolution trajectory and kernel density distribution within the first training round of the model, revealing the accelerating effect of the distillation mechanism on early feature polarization and gradient optimization; Figure 8 The following is a genome interaction feature analysis diagram based on GIA-Net provided for the embodiments of this application; wherein (a). Statistical distribution of MDIS scores of SNPs; (b). Cumulative SNP importance percentage of each chromosome; (c). Genome-wide interaction circular heatmap; (d). Local genetic interaction diagram of chromosome 8; Figure 9 Interpretability analysis of key variant sites provided for embodiments of this application; wherein (a) SNP whole-genome Manhattan plot; (b) SNP circular heatmap; (c) SHAP value analysis of the first 20 sites; Figure 10 The following diagram illustrates the evaluation of PRS prediction efficacy and the quantitative analysis of decision uncertainty based on BioMamba, provided for embodiments of this application; wherein (a) the PRS distribution characteristics and AUC of individuals in the test set; and (b) the correlation between PRS prediction values ​​and cognitive uncertainty under MC Dropout. Figure 11(a) A bar chart of GIA-Net ablation experimental results; (b) A bar chart of BioMamba ablation experimental results; (c) A feature ablation line chart with Top-K sites removed, provided for embodiments of this application; Figure 12 This is a structural diagram of an aquatic genetic trait prediction system based on multi-semantic feature fusion and knowledge distillation, provided in an embodiment of this application.

[0020] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0021] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0022] The collection, storage, use, processing, transmission, provision, and disclosure of relevant data and information in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0023] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0024] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0025] Aquaculture, as a core pillar of global high-quality protein supply, relies heavily on the breeding of superior varieties with durable disease resistance for its green and sustainable development. While genome-wide selection (GS) technology provides a key pathway to improve breeding efficiency, challenges such as high-dimensional feature redundancy, missing biological semantics in whole-genome data, and the enormous computational cost of traditional models in large-scale applications limit the implementation of precision breeding. To address these bottlenecks, this embodiment provides a method for predicting aquatic genetic traits based on multi-semantic feature fusion and knowledge distillation. This method can be based on... Figure 1 The implementation is carried out using a lightweight whole-genome prediction framework called BioMamba, which is not a simple application of a single model, but a progressive technology system consisting of three functionally synergistic modules: Geno-CLIP (Multi-Semantic Fusion Layer): Addressing the pain point of insufficient semantics in raw data, this layer utilizes cross-modal contrastive learning to map discrete sites to a continuous semantic feature space that integrates multi-source biological priors. This step aligns genetic signals with biological knowledge, constructing a feature foundation with deep implications for subsequent tasks.

[0026] Gene Interaction Mining Layer (GIA-Net): As a high-performance teacher network, it adopts a graph-sequence dual-branch architecture. The graph branch uses the associations between sites to resolve local linkage features, while the sequence branch uses the Mamba architecture to capture global long-range dependencies, aiming to deeply explore the nonlinear regulatory network behind disease resistance traits.

[0027] The high-efficiency inference layer (BioMamba) employs a state-space model with O(L) linear complexity. Through a three-layer collaborative distillation strategy involving response, features, and attention, it accurately inherits the gene interaction patterns inherent in GIA-Net. This model significantly reduces computational load while maintaining prediction accuracy highly consistent with expert models. Furthermore, BioMamba incorporates an attention-based importance scoring mechanism (MDIS), which explicitly traces and identifies core genetic loci determining disease resistance traits. Experimental results on datasets of 12 species, including rainbow trout and Atlantic salmon, demonstrate that this framework, with significantly reduced parameter counts, exhibits superior prediction robustness compared to traditional linear models and can accurately identify disease-causing candidate genes. This not only proves the effectiveness of multi-semantic fusion paths in enhancing model representation depth but also provides direct technical support for developing low-cost, high-performance domestically produced breeding chips.

[0028] Figure 1(a) After integrating prior information such as GWAS statistical priors and LD, cross-modal contrastive learning is performed through Geno-CLIP (contrastive language-image pre-training architecture). Discrete information is mapped to a continuous semantic feature space with biological connotations. (b) The teacher network GIA-Net adopts a "graph-sequence" dual-branch architecture for deep parsing of semantic embeddings. The graph branch is used to capture local linkages, and the sequence branch extracts global long-range dependencies. The dynamic fusion module explicitly simulates the interaction between the nonlinear control network and the superordinate. (c) Through a multi-level knowledge distillation strategy, the knowledge of the teacher network is transferred to the student network BioMamba.

[0029] Specifically, please combine Figure 2 As shown, the aquatic genetic trait prediction method based on multi-semantic feature fusion and knowledge distillation is implemented through the following steps S10-S50.

[0030] S10: Obtain whole-genome SNP data and corresponding disease resistance / susceptibility phenotype tags for aquatic animals.

[0031] The dataset used in step S10 of this embodiment covers SNP sequencing data of 12 real aquaculture species, including rainbow trout, Atlantic salmon, and oysters, as shown in Table 1. The phenotypic data mainly focuses on the binary classification trait of "disease resistance / susceptibility" under specific pathogen stress. This dataset has undergone a rigorous population genetics quality control (QC) process beforehand, including removing individuals and markers with a genotype loss rate greater than 10%, and filtering low-frequency variants based on the least allele frequency (MAF) to avoid statistical false positive associations under limited sample size conditions. After quality control, the genotypes are converted into a 0, 1, 2 numerical matrix based on a quantitative genetic additive effect model.

[0032] Table 1. Detailed information about the dataset

[0033] S20: Construct multi-semantic features and utilize a contrastive language-image pre-training architecture to map the original SNP data to a continuous semantic feature space that incorporates biological priors through cross-modal contrastive learning, thereby obtaining enhanced genomic feature embeddings.

[0034] In step S20, multi-semantic features are constructed based on the data obtained in step S10, and data enhancement is achieved by designing a feature semantic enhancement mechanism based on CLIP contrastive learning, thereby obtaining enhanced genomic feature embeddings.

[0035] In some embodiments, to address the issue of semantic missingness in the original SNP data, this embodiment designs Geno-CLIP (contrastive language-image pre-training architecture), such as... Figure 1As shown in (a), a six-dimensional genome semantic feature set was first constructed, and the one-dimensional sequence was reshaped into a two-dimensional feature tensor to adapt to the input of the visual encoder. The specific calculation formulas for each semantic feature are as follows: Genotypic semantic features (SNPVs) directly encode the original genetic variation of an individual. To capture the basal allele dosage effect, the SNPVs are... i Genotype of each SNP Normalize to the [0,1] interval: (1) in, For the first i The original genotype values ​​of each sample; Statistical Prior Semantic Feature GWASV: Addressing the common "complete separation" problem in binary disease resistance trait data, this embodiment abandons the traditional mixed linear model and employs Firth penalized logistic regression to obtain robust single-label effect estimates. For the first... j This embodiment aims to estimate the effect coefficient of 10 SNP loci. Assuming the sample i phenotypic probability Let be a nonlinear function with respect to the parameter to be determined: (2) in,, For the first i The conditional probability that a sample will exhibit a specific phenotype (such as disease resistance) given a given genotype. For the first n The true phenotypic label of each sample (e.g., 1 represents disease resistance, 0 represents susceptibility), This represents the SNP effect value. It is a natural exponential function. This is the intercept term of the logistic regression model. For the first n The genotype of each sample For the first c One principal component covariate, For the first c The effect coefficients of the principal component covariates, The total number of principal component covariates.

[0036] To eliminate the first-order bias in parameter estimation, a Firth-penalized log-likelihood function is constructed. This function finds the optimal solution by maximizing the joint probability of the observed data and introducing the Fisher information matrix as a penalty term. The specific formula is shown below: (3) in Let be the determinant of the Fisher information matrix. Using iterative algorithms such as the Newton-Raphson algorithm, we search for an unbiased estimator that maximizes the objective function. (4) in, Firth penalized log-likelihood function, It is a vector consisting of all parameters to be estimated in the model (including intercept term, principal component covariate coefficients, and SNP effect values).

[0037] Finally, after normalization Used as prior weights The corresponding statistical prior channels in the feature matrix: (5) This vector quantifies the statistical contribution of each locus to the disease resistance trait, providing crucial biological constraints for subsequent models.

[0038] Regulatory Hotspot Semantic Feature (REV) aims to identify hotspot SNPs located in potential regulatory regions. First, missing SNP data are imputed with mean, and reference allele frequencies are calculated. Then the frequency of the second allele was calculated. Set the 90th percentile as the threshold. If the product of the effect size and frequency exceeds a threshold, it is identified as a regulatory hotspot. (6) in, The 90th percentile threshold; This represents the frequency of the second allele. To examine allele frequencies; min is a minimum value function; It is a mean function; Genotype data of the training set after imputing missing values.

[0039] Gene density semantic feature (GDV): quantifies the local density of the genome. Defining a sliding window. The half-width is 50kbps, and the statistics are as follows: i The total number of variants within the neighborhood of each SNP And perform normalization: (7) in, For the first i The total number of variations within the neighborhood of each SNP For the firstk The physical location of each SNP For the first i The physical location of each SNP This represents the minimum total number of variants across all SNP neighborhoods. The maximum value of the total number of variants in the neighborhood of all SNPs. To prevent extremely small constants with a denominator of zero, The total number of SNPs on a chromosome or sequence.

[0040] Haplotype Structure Semantic Feature (HSV): Capturing population genetic characteristics under LD structure. First, the LD coefficient between any two loci is calculated. .like And physical distance Then they will be assigned to the same LD block. The channel value is defined as the absolute mean of the effect values ​​of all loci within the block containing the SNP: (8) in, For LD blocks, For the SNP index within the block, This is the SNP index currently being evaluated.

[0041] Quality Mask Semantic Feature (DQMV): Used as a binary filter to prevent noise interference. When the genotype at a certain locus is effective... Time-deleted sites are marked as 1, and missing or filled sites are marked as 0. (9) in, This encodes the original genotype at the i-th locus. The indicator function represents the semantic features of the quality mask when the genotype data is valid (i.e., a normal genotype with a value of 0, 1, or 2). The value is 1.

[0042] To provide semantic supervision signals, this embodiment designs dynamic prompt generation rules. Based on the sample phenotype... and characteristic statistics (such as controlled mean) Block proportion ), automatically generating structured text. For example, when At that time, the text was appended with the description "contains variants in highly active regulatory regions". The final text is composed of phenotypic and functional descriptions, forming a complete biological semantic tag. .

[0043] To adapt to the input of the ViT architecture, the above six semantic features are reconstructed into an RGB three-channel image through a weighted channel fusion strategy to complete feature alignment before cross-modal contrastive learning: (10) in, The pixel value for the red channel. The pixel value is for the green channel. The pixel value is for the blue channel. Genotype semantic features To statistically analyze prior semantic features, To regulate the semantic features of hotspots, For gene density semantic features, It is a semantic feature of haplotype structure. This refers to the semantic features of the quality mask.

[0044] Subsequently, comparative learning was performed using the CLIP architecture. A visual encoder extracted genomic features. Text encoder extracts text features The model aligns the two modalities in a shared feature space by optimizing a symmetric InfoNCE loss function. (11) in, The first extracted by the visual encoder i One genome modality feature, The first one extracted by the text encoder i A semantic modality feature, For temperature coefficient, This refers to the batch size.

[0045] This mechanism forces the model to learn the consistency between genetic variation and biological function, and the output feature vector This refers to a globally embedded system that has undergone semantic enhancement. This process corresponds to... Figure 1 (a) The semantic mapping stage from the original sequence input to the generation of the CLIP pre-trained feature matrix.

[0046] S30: Construct a teacher network with a dual-branch architecture, including a graph branch and a sequence branch. Embedded enhanced genomic features are input into the graph branch to obtain local structural features and the sequence branch to obtain a global sequence feature matrix. The local structural features and the global sequence feature matrix are then fused through a dynamic fusion module to generate the teacher network prediction results and site importance scores.

[0047] The data processing of the teacher network in step S30 is a core part of the method in this application. In general, based on the enhanced genomic feature embedding obtained in step S20, in the graph branch, local structural features are obtained by constructing and mining a genomic interaction network based on the graph structure. In the sequence branch, global sequence modeling based on the state space mechanism is used to obtain the global sequence feature matrix. The dynamic fusion module adopts a multi-scale feature fusion and global-local collaborative interaction mechanism to generate teacher network prediction results and site importance scores.

[0048] In some embodiments, the specific implementation of obtaining local structural features based on graph-structured genome interaction network construction and mining is as follows: To address the nonlinear interaction characteristics of functional variations across the entire genome, a strategy combining sliding windows and low-level lattice manipulation (LD) is employed to construct a genome heterogeneity map. Figure 1 Input for the branch in Figure (b). Set the sliding window size. 250kb, sliding step size It is 125kb. For the first [unit] in the figure... k Nodes Its covered SNP index set Defined as physical coordinates All sites falling within this window range: (12) in, This is an index of SNP sites.

[0049] Edge weights between nodes in the graph It is determined by both spatial physical proximity and genetic linkage strength, and the maximum of the two is taken to retain the most significant association, that is: (13) Where max is the maximum value function. For nodes and LD weights between; spatial proximity weights Depends on the physical coordinates of the node center and The distance, mapped by the Gaussian kernel function, is: (14) in, and These are the center physical coordinates of the two nodes.

[0050] LD weight Then the linkage correlation coefficient of the SNP pair Take the maximum value at time If the value is below a threshold, no connection is established. For any graph node...u Its initial eigenvector It is composed of the sum of the enhanced features of all SNPs within the coverage set, that is: (15) in, For the embedding of the i-th enhanced genomic feature, For nodes The set of SNP indexes included.

[0051] This indexing and aggregation process reduces the dimensionality of discretely distributed micro-variable signals and converges them into macro-block features, realizing the transformation from one-dimensional linear sequences to graph structures. After obtaining the initial features of block nodes, the GGMM module is used to mine local genetic patterns within blocks and suppress feature smoothing during deep network aggregation. First, a graph isomorphic network (GIN) is introduced to mine local genetic patterns within blocks. A multilayer perceptron (MLP) is used to perform injective transformations on the features of the central node and its neighboring nodes, and residual connections are combined to preserve the original sequence information, including its first-layer hidden features. Updated to: (16) in, For learnable scalar parameters; For layer normalization function; It is a multilayer perceptron; For nodes i The initial feature vector; For nodes i The set of neighboring nodes; For neighboring nodes j The initial eigenvectors.

[0052] Subsequently, considering the asymmetry of epistatic effects among different sites in the genome, a dynamic graph attention network (GATv2) was employed for high-order information interaction between nodes. i with neighbors j Dynamic attention coefficient The results were obtained through linear transformation of the spliced ​​features and the LeakyReLU activation function: (17) in, The weight matrix is ​​a learnable matrix. For nodes j The first layer of latent features This is a matrix transpose operation. This is a learnable attention weight vector.

[0053] Dynamic attention coefficient Weights obtained by Softmax normalization Then, weighted aggregation is performed on the neighborhood features, and residuals and layer normalization are superimposed again to obtain the final local explicit structural features. : (18) This mechanism overcomes the limitations of static attention in representing complex heterogeneous graphs. By dynamically adjusting the information transmission intensity between nodes, it accurately captures the pathogenic gene collaborative interaction network in the local physical space.

[0054] Given that graph neural networks are inherently permutation-invariant, they are prone to losing linear positional information of the genome on chromosomes during aggregation. To reconstruct this critical global order dependency, the local feature set output from the previous stage is reordered according to its corresponding physical coordinates to construct an ordered feature sequence. For the sorted number t The sequence is then processed by recursively scanning the input feature vectors of each node. Its internal hidden states are then analyzed. and the corrected node features The updated formula is as follows: (19) in, , For time-varying parameters, To output the projection matrix, This is the hidden state from the previous moment. For the sorted number t The input feature vector of each node. This process reconstructs the linear context of the chromosome, enabling the model to effectively capture long-range associations severed by the graph structure. Considering the significant differences in functional mechanisms among different regions of the genome, a single shared parameter network is insufficient to achieve optimal fitting of the entire genome; therefore, a hybrid expert module (MoE) is introduced at the end of the module. For any node feature... A Top-K gating mechanism incorporating standard normal Gaussian noise is introduced to calculate the matching confidence with each expert network: (20) in, The weight distribution of the gating network output. To retain the top K maximum values, For the learnable weight matrix of the gated network, It is Gaussian noise with a standard normal distribution.

[0055] Final local structural features The weighted sum of the MoE outputs is then superimposed with the residuals and normalized: (twenty one) in, For the first e The weighting of each expert. For expert network indexing, The total number of expert networks, For the first e The output of an expert network.

[0056] The above approach ensures that the model can dynamically invoke specific subsets of parameters for refined feature extraction targeting genomic blocks with different biological attributes. The final output feature matrix integrates local associations and global long-range dependencies, providing a high-quality structured representation for subsequent collaborative interaction mining.

[0057] In some embodiments, the specific implementation of obtaining the global sequence feature matrix based on state-space mechanism global sequence modeling is as follows: Local graph networks have limited receptive fields, and self-attention mechanisms suffer from quadratic computational complexity when processing very long whole-genome sequences. Therefore, in Figure 1 (b) Constructs a global sequence branch based on the Mamba architecture to perform full-context modeling of SNP sites across the entire genome with a linear complexity of O(L). This branch processes the genome as a one-dimensional continuous sequence. Its core structure includes an H3 (Hungry Hippo Head) architecture and a gated multilayer perceptron (Gated MLP), forming a dual-stream parallel network for feature extraction and information flow control. In the main feature branch, the sequence input first extracts the local physical associations between adjacent sites through one-dimensional convolution. After processing by the SiLU activation function, it enters the SSM operator. In this embodiment, the number of Mamba layers is set to 2, and the hidden layer feature dimension (… The state space dimension is uniformly set to 64. Set to 16, local one-dimensional convolution kernel size ( The parameter is set to 4. Combining continuous-time control theory with zero-order hold (ZOH) discretization technique, SSM dynamically generates time-varying parameters through a selective scanning mechanism. The discretized state transition matrix... With projection matrix The definition is as follows: (twenty two) in As the step size in the discretization of a continuous system, its function is to transform continuous-time parameters into discrete state transition matrices. and projection matrix This allows the model to process discrete SNP sequences recursively. It is the identity matrix. For time step, It is a continuous state transition matrix. It is a continuous projection matrix.

[0058] After discretization, the hidden state The recursive evolution and output calculation form is as follows: (twenty three) in, This represents the discrete hidden state at the previous time step. The input sequence features at the current time.

[0059] Through the aforementioned recursive evolution, the hidden states are driven to extract long-range information, establishing implicit regulatory dependencies across the entire genome. In parallel gating branches, input features are linearly projected to generate control signals. This signal is then multiplied element-wise with the output of the main branch SSM, forming a dynamic input gate similar to a Long Short-Term Memory (LSTM) network. This gating mechanism adaptively adjusts the pass rate of sequence features to suppress background noise and amplify the weights of key variant signals. Finally, the fused features from the two branches are passed through an output projection layer and residual connections to obtain the global sequence feature matrix. This serves as the input basis for subsequent collaborative analysis of cross-chromosomal distal mutations.

[0060] In some embodiments, the dynamic fusion module employs a multi-scale feature fusion and global-local collaborative interaction mechanism to generate teacher network prediction results and site importance scores. The specific implementation method is as follows: Entering GIA-Net ( Figure 1 Alignment and fusion are required before (b) in the middle. A hierarchical feature integration module (HCA) is constructed to concatenate the prior weights of GWAS with the SNP sequence features. The first three channels use one-hot encoding to orthogonally expand the three states of the genotype, while the fourth channel strictly aligns with the GWAS effect weights calculated by the Firth regression above, which are injected into the network as biological prior constraints. Subsequently, a multi-head cross-attention mechanism is used to introduce pre-trained multi-semantic features. In this calculation, local structural features are mapped to query vectors. Semantic features (i.e., the global sequence feature matrix) are mapped to keys. Sum The attention distribution is calculated as follows: (twenty four) in, For the output of each attention head, For attention weights, T This is a matrix transpose operation. is the feature dimension of the key vector.

[0061] The outputs of each attention head are linearly projected through a Feature Reconstruction (FR) layer, and then superimposed with residual connections and layer normalization. The final output of the HCA module is an enhanced feature matrix containing local variation signals and macroscopic environmental information. This serves as the input source for the bi-branch sequence. After the bi-branch feature extraction is completed, a crossover mechanism is introduced again to quantify the contribution of each genetic locus in order to accurately locate key disease resistance genes at SNP resolution. This mechanism will extract the local graph branch features. Let Q be the global sequence features. Perform cross-attention computation for K and V. The output of this interaction module evolves into two independent task branches: (1) Trait prediction: The weighted contextual representation generated by cross-attention calculation is directly used as a high-order feature stream and input into the downstream classifier to perform disease resistance trait prediction. This process utilizes global contextual information after full information interaction to ensure the accuracy of trait discrimination.

[0062] (2) Gene interaction interpretability: Extracting a weight vector covering the entire sequence length from the cross-attention operator. SNPs within the same LD block share the same block weight. It is difficult to distinguish between true causal mutations and non-functional hitchhiking variants. Therefore, a micro-feature response factor is introduced for site-specific correction. The enhanced features output by the HCA module are then used. As a specific benchmark, the first i Locus importance score of each SNP Defined as macro block weight Hadamard product with microscopic projection characteristics: (25) in, For the first i Locus importance score of each SNP , For a learnable projection matrix, For activation function, For macro-block weights, For Hadama accumulation, For the first i The enhanced feature matrix of each SNP. i Locus importance score of each SNP The independent contribution of each locus to the disease-resistant phenotype was precisely quantified. Finally, a genome-wide SNP-SNP interaction matrix was constructed based on the corrected SNP score vectors. M : (26) in, This represents the importance score of the locus.

[0063] SNP-SNP interaction matrix M elements in The quantification of the first i The locus and the first j The joint response intensity of each site in the feature space eliminates the numerical uniformity within the block, providing a high-resolution mathematical basis for analyzing finely detailed superordinate regulatory networks.

[0064] S40: Construct a lightweight student network; adopt a multi-level knowledge distillation strategy, use the prediction results generated by the teacher network and the site importance score to train the student network, and obtain the trained lightweight prediction model BioMamba.

[0065] In some embodiments, to reduce the computational complexity of whole-genome prediction while preserving the teacher network's ability to locate key genes, a lightweight prediction model based on SSM, BioMamba, was constructed as a lightweight student network. Please refer to [reference needed]. Figure 1 (c) This model discards the graph branch in the teacher network, retaining only the linearly complex Mamba structure as the feature extraction operator, and reducing the number of Mamba layers to 1 for lightweight implementation. Reduced to 8 Reduced to 2. The training process employs multi-level knowledge distillation, enabling the lightweight backbone to simulate the site saliency distribution captured by the teacher network during the training phase. BioMamba's total loss function. Defined as task supervision loss Response-level distillation loss and feature-level attention distillation loss Weighted combination: (27) in , To balance the hyperparameters of the component weights, this embodiment sets them to 0.3 and 10.0 respectively. Setting it to 0.3 aims to ensure the model primarily targets true disease resistance hard labels to guarantee basic classification accuracy, while introducing teacher soft labels as regularization to prevent overfitting; and Setting it to 10.0 is to amplify the extremely small mean squared error of attention (MSE) loss on a larger scale, so as to prevent its gradient from being overwhelmed by the cross-entropy loss, thereby ensuring the effective alignment of spatial features.

[0066] Task monitoring loss The binary cross-entropy (BCE) function is used to constrain the model's classification accuracy against disease phenotypes. (28) in, B For batch size, This is the true phenotypic label for the sample. To predict the probability distribution.

[0067] Response-level distillation loss Force students to output scores online using mean squared error (MSE). Approximating the output score of the teacher network , using migration prediction confidence: (29) in The temperature coefficient is set to 3.0 in this embodiment. Its function is to smooth the Logits distribution of the teacher and student network outputs, amplify the subtle probability differences between different categories, and thus make it easier for the student network to learn the inter-class relationships implied in the teacher network's decision-making.

[0068] Feature-level attention distillation loss in this embodiment The focus is on aligning the logic of key site locations by minimizing the site importance score vector between teacher and student models. Difference implementation: (30) in L For sequence length, and The first generation generated by the teacher network and the student network respectively i The importance score of each locus is used to constrain the lightweight model to identify key disease resistance loci by simulating the distribution of feature responses in the absence of graph branches.

[0069] Joint optimization of multi-level loss functions ensures that a lightweight backbone can establish site importance assessment criteria consistent with those of heavy-duty collaborative networks while reducing parameter size. This scheme achieves robust localization of disease-related important sites while meeting the requirements for inference of ultra-long genome-wide sequences in linear time.

[0070] S50: Use the trained lightweight prediction model BioMamba to predict traits from the genomic SNP data to be predicted, and output the classification results of disease resistance or susceptibility.

[0071] In some embodiments, to evaluate the predictive performance of the proposed model (the trained lightweight prediction model BioMamba) in the genome-wide disease resistance trait classification task, accuracy (ACC), precision, recall, F1 score, and Matthews correlation coefficient (MCC) are selected as evaluation metrics. The calculation of these metrics is based on the confusion matrix, and the parameters involved are defined as follows: true positive (TP), true negative (TN), false positive (FP), and false negative (FN), where the positive class corresponds to highly resistant samples, and the negative class corresponds to susceptibility.

[0072] Based on the above basic parameters, the calculation formulas for each indicator are defined as follows: Accuracy reflects the overall correctness of the model in classifying all disease-resistant and disease-susceptible samples: (31) Precision and recall are used to evaluate the accuracy and coverage of the model in predicting disease-resistant samples, respectively. (32) (33) The F1 score is the harmonic mean of precision and recall, used to comprehensively measure the model's recognition performance on disease-resistant samples. Considering the often unbalanced distribution of positive and negative samples in aquatic breeding data, MCC is introduced as a robustness evaluation index. MCC comprehensively considers the four classification results in the confusion matrix, and its calculation formula is as follows: Furthermore, the ROC curve and its area under the curve (AUC) were used to evaluate the model's discriminative ability at different thresholds. A value closer to 1 indicates higher discriminative efficacy of the model in identifying individual disease resistance. Through comprehensive analysis of these indicators, the performance and generalization level of the lightweight model after inheriting teacher network characteristics can be fully evaluated.

[0073] To quantitatively evaluate the predictive performance of the lightweight BioMamba model across different species and genetic backgrounds, this embodiment conducted comparative experiments on 12 independent aquatic datasets. The benchmark models covered traditional statistical genetics methods, representative deep learning models, state-of-the-art sequence modeling methods, and GIA-Net as a source of knowledge distillation.

[0074] This embodiment constructs an evaluation system containing six key indicators: ACC and AUC are used to measure overall discriminative power; Precision and Recall are used to reflect the balance between precision and recall; and F1 score and MCC are introduced as robustness evaluation criteria to address the common class imbalance in disease resistance data.

[0075] Figure 3 The performance distribution of each model across 12 evaluation datasets is presented. From a macroscopic perspective, GIA-Net and BioMamba, located at the bottom, exhibit a consistent dark blue color across all metrics, dominating the overall evaluation. In contrast, the other benchmark models above mostly appear in yellow or light green areas, representing poorer performance.

[0076] ACC and AUC reflect the overall discriminative efficiency of the model. Experimental results show that BioMamba's prediction accuracy significantly outperforms traditional statistical methods such as GBLUP, and is highly consistent with the performance of GIA-Net. In the area under the curve (AUC) metric, which excludes sample proportion interference, BioMamba achieved an optimal value of 0.986 on the Oyster-OsHV dataset, surpassing the 0.982 achieved by GIA-Net with explicit graph branches. This result verifies that the knowledge distillation strategy designed for whole-genome sequence features can effectively suppress overfitting in complex genetic backgrounds, enabling the lightweight student network to maintain a high level of discriminative accuracy at different classification thresholds.

[0077] In disease-resistant breeding practices, the balance between precision and recall directly affects breeding efficiency. Analysis shows that some models, such as GenNet, exhibit significant prediction skewness on the Trout-BCDW2 dataset. While their precision values ​​are acceptable, their recall values ​​are low, suggesting a high risk of underreporting of disease-resistant individuals. In contrast, BioMamba demonstrates excellent balance, achieving a recall of 0.991 on the Oyster-OsHV dataset, significantly better than GBLUP's 0.936 and GIA-Net's 0.985, ensuring the safety of breeding high-value disease-resistant parents.

[0078] To address the common problem of sample imbalance in aquatic data, F1 score and MCC provide more stringent evaluation benchmarks. In the F1 score, BioMamba's performance closely matches that of GIA-Net. Particularly on the Sparusaurata-Phdp dataset, which has the highest classification difficulty, the traditional GBLUP has an MCC of only 0.128, close to the level of random prediction, while BioMamba reaches 0.666, closely matching the performance upper limit of GIA-Net (0.686). Experimental results demonstrate that traditional linear models struggle to capture complex non-additive genetic effects, while BioMamba effectively extracts and integrates such non-linear implicit features through a distillation mechanism, exhibiting predictive robustness approaching that of heavyweight networks even under extreme data conditions.

[0079] To delve into the internal learning process of this model in the whole-genome prediction task, this embodiment uses visualization to examine the intermediate state of model training, focusing on the semantic alignment capability during the pre-training stage and the evolution of the feature space during the process.

[0080] To verify the semantic alignment capability of the functional view pre-trained model for biological prior concepts, this embodiment uses four datasets—Trout-SRS1, Trout-IPNV, Oyster-OsHV, and Carp-KHVD—to conduct Geno-CLIP biological semantic alignment analysis. The experiments evaluated the model's average probability of matching positive descriptive text across four dimensions: "Genomic Diversity," "Resistance Profile," "Gene Density," and "Haplotype Structure." Figure 5 In the "resistance trait" dimension, which is directly related to the phenotype, the probability of the disease-resistant group (red bars) matching the correct biological description was significantly higher than that of the susceptible group (blue bars) in all four datasets; while in the other three dimensions representing the global genetic background, the two groups showed similar probability distributions. This indicates that the pre-trained model can accurately capture and anchor high-order biological semantics directly related to the disease-resistant phenotype.

[0081] To further evaluate the effect of the aforementioned cross-modal mapping on the reshaping of the underlying data distribution, this embodiment compares the original SNP genotype feature space with the feature space distribution after Geno-CLIP enhancement of these four datasets. Figure 5 Here, the Silhouette Score is introduced to measure the quality of the feature space representation. This metric quantifies the clustering effect by comprehensively evaluating the intra-class compactness and inter-class separation of samples; the closer the value is to 1, the clearer the class boundaries. Experiments show that the four datasets in the original feature space ( Figure 5The silhouette coefficient in (a) is extremely low, with a Trout-SRS1 score of only -0.002, indicating that the original genotype is insufficient to effectively distinguish phenotypic differences. However, in the enhanced feature space (… Figure 5 In (b)), the samples exhibit a clear manifold structure, with both groups effectively pushed to opposite sides of the feature space. Quantitative analysis shows that the silhouette coefficients of all four datasets achieve orders-of-magnitude improvements, such as Oyster-OsHV reaching 0.272 and Trout-IPNV reaching 0.225. The results demonstrate that the introduction of functional semantics effectively filters out a large amount of genetic background noise in the early stages of feature extraction, reshapes the data distribution through cross-modal mapping, and provides a high signal-to-noise ratio representational foundation for accurate modeling of downstream dual-branch networks.

[0082] The student network inherits the discriminative ability of the teacher network through knowledge distillation. To intuitively analyze the dynamic evolution of the feature space during this knowledge transfer process, this embodiment uses the Trout-SRS1 dataset and employs t-SNE dimensionality reduction technology to visualize and track the deep feature distribution of samples at different training epochs. In the early stages of training, such as... Figure 6 In (a), there is still a significant overlap between high- and low-disease-resistance samples in the feature space, and the inter-class classification boundaries have not yet been established. With the continued intervention of soft-label distillation, such as... Figure 6 In (b) and (c), the feature vectors of the two classes of samples undergo significant stretching and peeling along a specific manifold direction. Towards the end of training, as... Figure 6 In the middle (d), the feature space of BioMamba has been highly structured, and the disease-resistant and susceptible groups are completely pushed to the two poles of the space, forming independent and compact clusters.

[0083] Furthermore, the accelerating effect of the distillation mechanism on early model convergence is objectively reflected in the evolution of feature representation. This embodiment uses the micro-trajectory of the Trout-SRS1 dataset during the first training epoch, from 10% to 100% iterations. Figure 7 This demonstrates the directional movement pattern of the average eigenvectors of high-resistance and low-resistance populations. Figure 7 The background kernel density contour lines characterize the distribution range of samples in the latent space. The results show that even during the random initialization of network parameters, the representation paths of the two groups do not exhibit random wandering, but rather rapidly diverge towards the extremes of the feature space. This representation evolution phenomenon confirms that the soft distribution of the teacher network output effectively transmits structured category priors, guiding the student network to accurately lock onto the effective feature subspace of associated phenotypic differences from the very beginning of training.

[0084] In disease-resistant breeding applications, besides focusing on the overall predictive accuracy of the model, identifying and locating key SNP sites driving phenotypic variations is also of significant practical value. This example uses the Trout-SRS1 dataset to analyze genome-wide variation sites using the constructed model.

[0085] The contribution weights and interaction strengths of each genetic locus are directly extracted through GIA-Net's built-in collaborative attention mechanism and feature projection layer. The MDIS score is derived from the Hadamard product of macroscopic block weights generated by multi-head cross-attention and deep locus features after nonlinear projection. This mechanism ensures that the model can accurately capture the independent contribution of each unit point while preserving the global context. Based on this score vector, the model constructs a genome-wide interaction matrix by performing tensor outer product operations, aiming to quantify the joint response strength between loci in the latent space, thereby transforming high-dimensional nonlinear features into a biologically interpretable numerical distribution.

[0086] Figure 8 The MDIS scores in (a) showed a significant concentration around zero, with only a very small number of loci located in the top 5% of high-scoring intervals. This quantitatively revealed the genetic sparsity of the disease resistance trait, i.e., the core variants were concentrated in a few key functional loci. Figure 8 In section (b), the cumulative contribution of each chromosome was further ranked, with chromosome 8 showing the highest cumulative contribution, indicating its central role was significantly superior to other linkage groups. Further in-depth analysis of this region was conducted. Figure 8 In (c), the spiral coordinate system was used to characterize the spatial distribution of features within chromosome 8. High-scoring nodes showed obvious clustering characteristics in physical location, and long-range connections across layers proved that the model captured implicit functional couplings beyond the linear distance of the genome. Figure 8 (c) shows the importance of each gene and the information of strongly correlated interactions by screening based on thresholds through a circular heatmap. Figure 8 (d) A rigorous dynamic thresholding of the interaction matrix using a linear view clearly reveals the star-shaped topological hub nodes near 26.00 Mb, demonstrating the pleiotropic regulatory role of key loci. Multiple long-range arcs spanning over 30 Mb exceed the decay range of conventional linkage disequilibrium, providing direct computational evidence for long-distance epistatic effects and potential chromatin circularization mechanisms. These results demonstrate that GIA-Net can not only accurately locate key functional loci but also uncover complex genetic regulatory networks spanning long physical distances through synergistic weighting.

[0087] To further determine the importance of core sites and BioMamba's knowledge inheritance capabilities, this embodiment extracts the attention weights assigned to each SNP by its internal selection mechanism. Through nonlinear scaling processing, Figure 9The genome-wide attention Manhattan plot shown in (a) depicts the relative discriminative strength of loci. A higher peak on the vertical axis indicates a greater contribution of the locus to the model's decision. Within the set threshold of the first 0.5%, the model identified multiple steep signal peaks on chromosomes 4, 7, 10, 11, and 18, revealing core causal mutations with extremely high independent weights. This distribution pattern logically complements the high cumulative score in the chromosome 8 region, confirming that BioMamba can effectively remove redundant interference from linkage disequilibrium and accurately pinpoint major functional loci.

[0088] Figure 9 The circular genome heatmap shown in (b) spatially reproduces the aforementioned signals. The extremely bright yellow to bright pink regions in the heatmap are highly spatially coupled with the discrete peaks in the Manhattan plot. Cross-validation combining point and surface data eliminates the interference of accidental false positives. To further reveal the driving forces of the traits, Figure 8 c. Using SHAP values, the specific genetic effects of these core loci were analyzed. Taking the top-ranked locus, Affx.88961105, as an example, the red scatter plots carrying high coding values ​​significantly clustered in the negative interval, demonstrating a clear disease resistance effect; while the blue scatter plots carrying low coding values ​​clustered in the positive interval, indicating an increased risk of disease. This clear red-blue, left-right opposition distribution pattern analyzed the positive and negative driving mechanisms of core loci at the microscopic level, and numerically reconstructed the classic allele additive dosage effect in genetics, fully demonstrating that BioMamba has the application value of analyzing complex genetic logic and guiding breeding screening.

[0089] In genomic selection breeding, a single point estimate of probability is insufficient to fully reflect the reliability of model predictions. If the model misjudges individuals carrying complex recombination variations, incorrectly selecting susceptible individuals as disease-resistant parents, it poses a serious risk to breeding practices. Therefore, quantifying the cognitive uncertainty of model predictions and identifying high-risk samples is a crucial step in improving the safety of whole-genome predictions. After analyzing the genetic effects of core loci, this embodiment evaluates the performance of the BioMamba model on the Trout-SRS1 dataset to further assess its robustness in practical breeding decisions. Figure 10Figure (a) shows the distribution of the polygenic risk score (PRS) of individuals in the test set and its overall classification performance. The results show that the PRS distribution exhibits a significant bipolar characteristic: the predicted values ​​of the vast majority of samples are highly compressed within the endpoints of 0 (resistance) or 1 (susceptibility), while very few ambiguous samples are located near the 0.5 decision boundary. This highly discriminative bimodal distribution directly boosts the model's overall classification efficiency, achieving an AUC of 0.9549. This indicates that BioMamba possesses extremely high classification confidence and a clear decision boundary, providing a deterministic mathematical basis for defining robust selection thresholds in breeding practice. However, a single point estimate probability cannot completely eliminate the risk of misjudgment due to rare recombination or complex heterozygous backgrounds. Therefore, this embodiment introduces MC Dropout in the inference stage, measuring cognitive uncertainty by quantifying the variance of 30 random forward propagation results. The correlation between the obtained PRS and uncertainty is as follows: Figure 10 As shown in (b), the data distribution exhibits a clear inverted U-shaped trend. Individuals in the two extreme high-confidence regions correspond to extremely low uncertainty indices, indicating that the model can maintain highly consistent outputs for genotypes with significant disease resistance or susceptibility. However, individuals near the 0.5 decision boundary, due to the complexity of their genetic background, cause oscillations in the selective paths within the model, resulting in significantly increased prediction volatility and forming the peak region of uncertainty. Through integration... Figure 10 The predictive tendency of (a) and Figure 10 Based on the confidence level in (b), this embodiment constructs a two-dimensional screening system for disease-resistant germplasm. Breeders can not only initially identify disease-resistant populations based on low PRS values, but also eliminate low-confidence individuals located in high-fluctuation zones by setting an uncertainty threshold on the vertical axis. This joint filtering mechanism effectively overcomes the limitations of traditional point estimation, excluding potential genotypic variation risks from the core breeding population, and providing rigorous computational evidence to ensure the phenotypic stability of the Trout-SRS1 breeding population.

[0090] To verify the necessity of the multi-dimensional feature co-distillation strategy and its innovative components, this embodiment performed ablation experiments and feature sensitivity tests on the Trout-SRS1 dataset. By modularly decomposing the architectural components of the teacher network GIA-Net and the distillation strategy of the student network BioMamba, the actual gain of each dimension of information on prediction performance was quantified.

[0091] The robustness of the GIA-Net teacher network architecture stems from the synergistic gains of its core components. Table 2 records the performance benchmarks under different module absences, while... Figure 11 The middle (a) chart visually illustrates the different trends of each indicator after it loses specific support.

[0092] Table 2. Results of GIA-Net ablation experiments

[0093] It can be seen that the removal of CLIP pre-training caused a sharp drop in the Recall metric (down to 0.642), which confirms that biological semantic priors are key clues for the model to identify rare disease-resistant samples. Meanwhile, the loss of sequence and graph branches induced a comprehensive performance degradation, demonstrating the irreplaceable role of integrating multidimensional heterogeneous features in characterizing complex causal variations. After establishing the optimal teacher architecture, this embodiment further examined the accuracy of the student network BioMamba in inheriting genetic knowledge, and the effectiveness of the relationship between semantic features and graph structure in knowledge transfer. The evolution path of the distillation strategy is recorded in Table 3, while the effect of each loss term on the correction of prediction bias is shown in... Figure 11 The concrete representation is shown in (b).

[0094] Table 3 Results of BioMamba ablation experiments

[0095] Experimental results show that, under the condition of knowledge distillation (KD), the prediction bias of the lightweight model is effectively corrected, significantly improving the precision from 0.593 to 0.906, achieving a biological balance between precision and recall. When the teacher network CLIP (lacking CLIP pre-training) is used, the recall drops sharply to only 0.280. This strongly demonstrates that biological semantic priors are the core clues for the model to identify rare disease-resistant samples. When the teacher network Graph (lacking graph structure branches), that is, when gene interaction information is missing, the F1 score drops to 0.584, verifying the importance of local LD ​​information for finely characterizing decision boundaries and reducing misjudgments. Finally, regardless of removing... (Attention weight feature distillation) or (Logits soft-label distillation) will cause a significant drop in AUC and MCC. (Removal) Subsequently, Recall decreased from 0.828 to 0.650, indicating that aligning the attention weights of the teacher network significantly improved the student network's sensitivity to functional loci localization. (Missing information) This leads to ambiguity in the inter-class discriminative boundaries, indicating that the inter-class correlation information provided by soft labels is fundamental to maintaining the accuracy of prediction confidence. In summary, BioMamba's superior performance does not stem from the accumulation of single modules, but rather from the deep fusion and synergistic effect of multi-source knowledge under a multi-dimensional distillation strategy. By introducing a distillation mechanism, the model successfully absorbed the inter-class association priors from the teacher network, significantly increasing precision to 0.906, achieving a better biological trade-off between precision and recall.

[0096] To confirm whether the SNP sites located by BioMamba truly carry core genetic information, this embodiment concludes with a feature removal experiment. As the top K sites with the highest importance scores were sequentially removed, the model's predictive performance exhibited a significant non-linear decline. This dynamic process is as follows: Figure 11 As shown in (c), after removing the top 50 high-weight loci, the AUC rapidly decreased from 0.958 to 0.898; when the number of removals reached 500, the performance further bottomed out at 0.838. This steep decline in the early stage strongly demonstrates that the loci selected by BioMamba do indeed integrate the key genetic signals that determine the phenotype, rather than background noise. However, as the number of removals continues to increase, the slope of the curve gradually flattens, and even after removing approximately 40% of the total number of loci, the model still maintains a significantly higher predictive ability than random. This phenomenon, where the prediction results change with the loss of core loci, profoundly reflects the complex genetic architecture of disease resistance traits, namely, that traits are driven by a few major genes and a large number of widely distributed minor variations. BioMamba's accurate capture of this complex polygenic effect has not only been rigorously validated on the Trout-SRS1 dataset, but also demonstrates its strong robustness and broad universality in handling prediction tasks in other species or across populations with similar polygenic genetic backgrounds.

[0097] Based on the above experimental results, it is evident that this application proposes a prediction method based on multi-semantic features and knowledge distillation to address the problems of semantic feature loss, high-dimensional feature redundancy, difficulty in long sequence modeling, and lack of model interpretability in genome-wide prediction. Experimental results on datasets such as Trout-SRS1 and Oyster-OsHV demonstrate that this method can effectively overcome the interference of genetic background noise. Through the Geno-CLIP pre-training module, the model achieves cross-modal alignment between genome sequences and higher-order biological semantics, effectively filtering genetic background noise in the early stages of feature extraction and reshaping a high signal-to-noise ratio feature base. GIA-Net leverages the advantages of graph neural networks and attention mechanisms to locate epistatic effects between loci across the entire genome. Through MDIS index analysis, the model not only identifies core candidate loci on key chromosomes but also reconstructs the complex network formed by nonlinear associations between loci, providing a visualization tool for the genetic analysis of complex traits. Finally, knowledge distillation significantly reduces the parameter count and inference overhead of BioMamba, and the core evaluation metrics are superior to existing benchmark models. Through internal attention mechanisms combined with SHAP analysis, the core SNP loci driving phenotypic variations are accurately located. Furthermore, the ablation experiment verified the necessity of the collaborative design of pre-training, graph space architecture and knowledge distillation mechanism, while the MC Dropout mechanism also provides a quantitative means for identifying difficult samples and avoiding the risk of misjudgment in actual breeding.

[0098] In summary, the method presented in this application effectively balances the lightweight nature of the network and its biological interpretability while ensuring the accuracy of trait prediction. This method not only improves the computational efficiency and security of genome-wide selection but also serves as a reliable tool for discovering key mutation sites, demonstrating high application value in practical disease-resistant breeding of plants and animals.

[0099] This application also provides an aquatic genetic trait prediction system based on multi-semantic feature fusion and knowledge distillation, used to implement the methods described in any of the above embodiments, such as... Figure 12 As shown, the system includes: The data acquisition module 1201 is configured to acquire whole genome SNP data of aquatic animals and corresponding disease resistance / susceptibility phenotype tags; The data augmentation module 1202 is configured to construct multi-semantic features and utilize a contrastive language-image pre-training architecture to map the original SNP data to a continuous semantic feature space that incorporates biological priors through cross-modal contrastive learning, thereby obtaining enhanced genomic feature embeddings. The teacher network construction module 1203 is configured to construct a teacher network, which adopts a dual-branch architecture, including a graph branch and a sequence branch. The enhanced genomic features are embedded and input into the graph branch to obtain local structural features, and into the sequence branch to obtain a global sequence feature matrix. The local structural features and the global sequence feature matrix are fused through a dynamic fusion module to generate teacher network prediction results and site importance scores. The student network training module 1204 is configured to construct a lightweight student network; using a multi-level knowledge distillation strategy, the student network is trained using the prediction results and site importance scores generated by the teacher network to obtain the trained lightweight prediction model BioMamba. The classification prediction module 1205 is configured to use the trained lightweight prediction model BioMamba to predict traits of the genomic SNP data to be predicted and output the classification result of disease resistance or susceptibility.

[0100] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting aquatic genetic traits based on multi-semantic feature fusion and knowledge distillation, characterized in that, The method includes: Obtain whole-genome SNP data and corresponding disease resistance / susceptibility phenotype tags for aquatic animals; We construct multi-semantic features and utilize a contrastive language-image pre-training architecture to map the original SNP data to a continuous semantic feature space that incorporates biological priors through cross-modal contrastive learning, thereby obtaining enhanced genomic feature embeddings. A teacher network is constructed, which adopts a dual-branch architecture, including a graph branch and a sequence branch. The enhanced genomic features are embedded and input into the graph branch to obtain local structural features and the sequence branch to obtain a global sequence feature matrix. The local structural features and the global sequence feature matrix are then fused through a dynamic fusion module to generate teacher network prediction results and site importance scores. A lightweight student network is constructed; a multi-level knowledge distillation strategy is adopted, and the student network is trained using the prediction results generated by the teacher network and the site importance scores to obtain the trained lightweight prediction model BioMamba. The trained lightweight prediction model BioMamba is used to predict traits from the genomic SNP data to be predicted, and outputs the classification results of disease resistance or susceptibility.

2. The method according to claim 1, characterized in that, The multi-semantic features include six dimensions of genomic semantic features, namely, genotype semantic features, statistical prior semantic features, regulatory hotspot semantic features, gene density semantic features, haplotype structure semantic features, and quality mask semantic features. Among them, genotype semantic features The calculation formula is as follows: (1) in, For the first i The original genotype values ​​of each sample; Statistical prior semantic features The calculation method is as follows: Determine the sample n phenotypic probability Let be a nonlinear function with respect to the parameter to be determined: (2) in, For the first i The conditional probability that a sample will exhibit a specific phenotype given a given genotype. For the first n The true phenotypic label of each sample This represents the SNP effect value. It is a natural exponential function. This is the intercept term of the logistic regression model. For the first n The genotype of each sample For the first c One principal component covariate, For the first c The effect coefficients of the principal component covariates, The total number of principal component covariates; Construct the Firth penalized log-likelihood function : (3) in, This is the determinant of the Fisher information matrix; The unbiased estimator that makes the function reach its global maximum is obtained by solving the Newton-Raphson iterative algorithm. : (4) in, Firth penalized log-likelihood function, This is a vector consisting of all the parameters to be estimated in the model; Normalized SNP effect values As prior weights The statistical prior semantic features are calculated using the following formula, corresponding to the statistical prior channels in the feature matrix. : (5) Regulation of hotspot semantic features The calculation formula is as follows: (6) in, The 90th percentile threshold; This represents the frequency of the second allele. To examine allele frequencies; min is a minimum value function; It is a mean function; Genotype data of the training set after imputing missing values; Semantic features of gene density The calculation formula is as follows: (7) in, For the first i The total number of variations within the neighborhood of each SNP For the first k The physical location of each SNP For the first i The physical location of each SNP This represents the minimum total number of variants across all SNP neighborhoods. The maximum value of the total number of variants in the neighborhood of all SNPs. To prevent extremely small constants with a denominator of zero, The total number of SNPs on a chromosome or sequence; Haplotype structural semantic features The calculation formula is as follows: (8) in, For LD blocks, For the SNP index within the block, For the currently evaluated SNP index; Quality mask semantic features The calculation formula is as follows: (9) in, This encodes the original genotype at the i-th locus. The indicator function represents the semantic features of the quality mask when the genotype data is valid (i.e., a normal genotype with a value of 0, 1, or 2). The value is 1.

3. The method according to claim 2, characterized in that, Methods for contrastive learning using a contrastive language-image pre-training architecture include: The six-dimensional genomic semantic features are reconstructed into an RGB three-channel image using the following formula: (10) in, The pixel value for the red channel. The pixel value is for the green channel. The pixel value is for the blue channel. Genotype semantic features To statistically analyze prior semantic features, To regulate the semantic features of hotspots, For gene density semantic features, It is a semantic feature of haplotype structure. For quality mask semantic features; By optimizing the symmetric InfoNCE loss function Align the two modalities in a shared feature space, where Represented as: (11) in, The first extracted by the visual encoder i One genome modality feature, The first one extracted by the text encoder i A semantic modality feature, For temperature coefficient, This refers to the batch size.

4. The method according to claim 1, characterized in that, Methods for embedding the enhanced genomic features into the graph branch to obtain local structural features include: A strategy combining sliding window and linkage disequilibrium is used to construct a genome heterogeneity map, wherein the first... k Nodes SNP index set Represented as: (12) in, For SNP site indexing, The sliding step size, For physical coordinates, This refers to the size of the sliding window; Edge weights between nodes Calculated using the following formula: (13) Where max is the maximum value function. For nodes and The LD weights between them; The spatial proximity weight is calculated using the following formula: (14) in, and The center physical coordinates of the two nodes; Initial feature vector of a node Calculated using the following formula: (15) in, For the embedding of the i-th enhanced genomic feature, For nodes The set of SNP indexes included; Based on the aforementioned genomic heterogeneity map, local genetic patterns are mined using graph isomorphism networks and dynamic graph attention networks, and a Mamba module is introduced to reconstruct linear order dependencies. The local structural features are then output through a hybrid expert module.

5. The method according to claim 4, characterized in that, Based on the aforementioned genomic heterogeneity map, local genetic patterns are mined using graph isomorphism networks and dynamic graph attention networks. A Mamba module is introduced to reconstruct linear order dependencies, and the local structural features are output through a hybrid expert module. Specifically, these features include… In the graph isomorphic network, the first layer hidden features The update formula is: (16) in, For learnable scalar parameters; For layer normalization function; It is a multilayer perceptron; For nodes i The initial feature vector; For nodes i The set of neighboring nodes; Neighboring nodes j The initial feature vector; In the dynamic graph attention network, nodes i with neighbors j Dynamic attention coefficient The calculation formula is: (17) in, The weight matrix is ​​a learnable matrix. For nodes j The first layer of latent features This is a matrix transpose operation. This is a learnable attention weight vector; Local explicit structural features after Softmax normalization The update formula is as follows: (18) in, Dynamic attention coefficient Weights obtained after Softmax normalization; After importing the Mamba module, the hidden state is... and the corrected node features The update formula is as follows: (19) in, , For time-varying parameters, To output the projection matrix, This is the hidden state from the previous moment. For the sorted number t The input feature vector of each node; The hybrid expert module has features for any node. A Top-K gating mechanism incorporating standard normal Gaussian noise is introduced to calculate the matching confidence with each expert network: (20) in, The weight distribution of the gating network output. To retain the top K maximum values, For the learnable weight matrix of the gated network, The noise is a standard normally distributed Gaussian noise. Local structural features are calculated using the following formula. : (21) in, For the first e The weighting of each expert. For expert network indexing, The total number of expert networks, For the first e The output of an expert network.

6. The method according to claim 1, characterized in that, Methods for embedding the enhanced genomic features into the sequence branch to obtain a global sequence feature matrix include: The enhanced genomic features are embedded and processed into a one-dimensional continuous sequence input sequence branch; The sequence branches adopt the Mamba architecture, including parallel feature main branches and gated branches; In the main feature branch, the input sequence extracts the local physical associations between adjacent sites through one-dimensional convolution, and after being processed by the SiLU activation function, it enters the state space model operator. In the state-space model operator, time-varying parameters are dynamically generated through a selective scan mechanism, wherein the discretized state transition matrix... With projection matrix The calculation formula is: (22) in, The step size is used for discretizing continuous systems; its function is to transform continuous-time parameters into discrete state transition matrices. and projection matrix , It is the identity matrix. For time step, It is a continuous state transition matrix. It is a continuous projection matrix; After discretization, the hidden state The recursive evolution and output calculation formulas are shown below: (23) in, This represents the discrete hidden state at the previous time step. The input sequence features at the current time step; In the gated branch, the input features are linearly projected to generate a control signal; The control signal is multiplied element-wise with the output of the main branch SSM operator to obtain the gated modulation features. The gated and modulated features are processed sequentially through an output projection layer and a residual connection. The residual connection adds the output of the output projection layer to the original input of the sequence branch to obtain the global sequence feature matrix.

7. The method according to claim 1, characterized in that, The local structural features and global sequence feature matrix are fused by a dynamic fusion module to generate teacher network prediction results and site importance scores, specifically including: Map local structural features to query vectors The global sequence feature matrix is ​​mapped to the key. Sum The output of each attention head is calculated through a multi-head cross-attention mechanism; wherein, the calculation formula of the multi-head cross-attention mechanism is: (24) in, For the output of each attention head, For attention weights, T This is a matrix transpose operation. The feature dimension of the key vector; The outputs of each attention head are linearly projected through a feature reconstruction layer (FR), and then superimposed with residual connections and layer normalization to obtain an enhanced feature matrix containing local variation signals and macroscopic environmental information. ; Enhance the feature matrix The inputs are respectively fed into a first task branch and a second task branch; wherein, the first task branch is used to process the enhanced feature matrix. The downstream classifier performs disease resistance trait prediction and obtains the predicted disease resistance trait as the teacher network prediction result; the second task branch is used to perform the prediction based on the enhanced feature matrix. The importance score of a site is calculated using the following formula: (25) in, For the first i Locus importance score of each SNP , For a learnable projection matrix, For activation function, For macro-block weights, For Hadama accumulation, For the first i Enhanced feature matrix of SNPs.

8. The method according to claim 1 or 7, characterized in that, After obtaining the site importance score, the method further includes constructing a genome-wide SNP-SNP interaction matrix based on the site importance score. M : (26) in, This represents the importance score of the locus.

9. The method according to claim 1, characterized in that, The total loss function used in the aforementioned multi-level knowledge distillation strategy Represented as: (27) in, , To balance the hyperparameters of the component weights, For the loss of mission supervision, For response stage distillation loss, Characteristic stage distillation loss; The task supervision loss Calculated using the following formula: (28) in, B For batch size, This is the true phenotypic label for the sample. To predict the probability distribution; The response-level distillation loss Calculated using the following formula: (29) in, Temperature coefficient; Output scores to students online; Scoring of the output of the teacher network; The feature-level attention distillation loss Calculated using the following formula: (30) in, For sequence length, and The first generation generated by the teacher network and the student network respectively i Importance score of each site.

10. A system for predicting aquatic genetic traits based on multi-semantic feature fusion and knowledge distillation, characterized in that, The system for implementing the aquatic genetic trait prediction method based on multi-semantic feature fusion and knowledge distillation as described in any one of claims 1-9, comprises: The data acquisition module is configured to acquire whole-genome SNP data of aquatic animals and corresponding disease resistance / susceptibility phenotype tags; The data augmentation module is configured to construct multi-semantic features and utilize a contrastive language-image pre-training architecture to map the original SNP data to a continuous semantic feature space that incorporates biological priors through cross-modal contrastive learning, thereby obtaining enhanced genomic feature embeddings. The teacher network construction module is configured to construct a teacher network, which adopts a dual-branch architecture, including a graph branch and a sequence branch. The enhanced genomic features are embedded and input into the graph branch to obtain local structural features, and into the sequence branch to obtain a global sequence feature matrix. The local structural features and the global sequence feature matrix are then fused through a dynamic fusion module to generate teacher network prediction results and site importance scores. The student network training module is configured to construct a lightweight student network; a multi-level knowledge distillation strategy is adopted to train the student network using the prediction results and site importance scores generated by the teacher network, resulting in the trained lightweight prediction model BioMamba. The classification prediction module is configured to use the trained lightweight prediction model BioMamba to predict traits in the genomic SNP data to be predicted, and output the classification result of disease resistance or susceptibility.

Images