A High-Throughput Intelligent Screening Method for Multi-Track Ginseng Breeding Materials Based on Deep Learning

Abstract

This invention relates to the field of intelligent breeding technology, specifically disclosing a high-throughput intelligent screening method for multi-trait ginseng breeding materials based on deep learning. The method includes breeding data acquisition, high-throughput acquisition of hyperspectral phenotypes, construction of a multi-trait prediction model, model interpretability analysis, comprehensive screening of multiple traits, and verification of screening results. This scheme utilizes UAV hyperspectral remote sensing technology to achieve rapid and non-destructive determination of saponin content, significantly improving the throughput of phenotypic data acquisition. A multi-task learning architecture and deep neural networks are used to construct a multi-trait prediction model. By sharing representation layers, genetic associations between traits are captured, while retaining the specific information of each trait, achieving collaborative and accurate prediction of multiple traits. Through integrated gradient and attention weight analysis, key marker sites are accurately located, and their biological significance is revealed by functional annotation, enhancing the model's transparency and credibility.

Description

Technical Field

[0001] This invention relates to the field of intelligent breeding technology, specifically to a high-throughput intelligent screening method for multi-trait ginseng breeding materials based on deep learning. Background Technology

[0002] Ginseng, a traditional and precious Chinese medicinal herb, possesses active ingredients, ginsenosides, which exhibit significant pharmacological effects in anti-inflammation, anti-oxidation, anti-tumor activity, and immunomodulation. With the accelerating internationalization of traditional Chinese medicine and the growing demand in the health industry, the need for high-quality new ginseng varieties is becoming increasingly urgent. Traditional ginseng breeding is characterized by long cycles and low efficiency, relying primarily on breeders' experience and field phenotypic observations, which fails to meet the precision and efficiency requirements of modern seed industry. Furthermore, existing deep learning methods lack transparency in their decision-making processes, making it difficult to reveal key marker sites and their biological significance, thus limiting their practical application in breeding decisions. Moreover, existing genomic selection methods in ginseng breeding typically construct predictive models for single traits, neglecting the intrinsic connections and synergistic effects between traits, making it difficult to meet the needs of comprehensive improvement of multiple traits in breeding practice. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a high-throughput intelligent screening method for multi-trait ginseng breeding materials based on deep learning. Addressing the problems of low efficiency, high destructiveness, and the neglect of trait associations in traditional phenotypic identification methods, this solution utilizes UAV hyperspectral remote sensing technology to achieve rapid and non-destructive determination of saponin content, significantly increasing the throughput of phenotypic data acquisition. A multi-task learning architecture and deep neural networks are used to construct a multi-trait prediction model. By sharing the representation layer, genetic associations between traits are captured, while retaining the specific information of each trait, achieving collaborative and accurate prediction of multiple traits. Addressing the pain point of poor interpretability in deep learning models, this solution integrates gradient and... Dual attention weight analysis accurately locates key marker sites and reveals their biological significance by combining functional annotation, enhancing the transparency and credibility of the model. This provides theoretical guidance for molecular marker-assisted breeding and ensures good generalization performance on independent test sets, providing reliable assurance for practical breeding applications. Addressing the challenge of multi-trait comprehensive screening, this scheme combines multi-trait prediction results with multi-objective optimization, scientifically balancing the trade-offs between various breeding objectives. This enables precise grading and intelligent screening of breeding materials, significantly improving the scientific rigor and efficiency of breeding decisions, establishing a complete breeding closed loop, and ensuring the technical solution can be effectively applied in practical breeding work, generating continuous technical benefits.

[0004] The technical solution adopted in this invention is as follows: The high-throughput intelligent screening method for multi-trait ginseng breeding materials based on deep learning provided by this invention includes the following steps:

[0005] Step S1: Breeding data collection. Collect ginseng germplasm resources to construct breeding material populations. Perform genotyping and phenotypic determination on each breeding material to obtain genotypic and phenotypic data, in the following format:

[0006] ;

[0007] ;

[0008] In the formula, Representing genotype data in matrix form, Index representing breeding materials, An index representing a SNP site. Indicates the first The breeding material was in the first The genotype encoding of each SNP locus Representing tabular data in matrix form, Represents a trait index. Indicates the first The breeding material was in the first Standardized phenotypic values ​​of individual traits;

[0009] Step S2: High-throughput acquisition of hyperspectral phenotypes. Hyperspectral images of ginseng in the experimental field are collected using a drone equipped with a hyperspectral imaging system. Low-dimensional spectral feature vectors are obtained through preprocessing and feature extraction. A quantitative correlation model between spectral features and quality traits is established to augment the phenotypic data in a high-throughput manner.

[0010] Step S3: Construction of multi-trait prediction model. A multi-task learning architecture and deep neural network are used to construct a multi-trait prediction model. Based on genotype data and low-dimensional spectral feature vectors, the hyperparameters of the multi-trait prediction model are optimized using ten-fold cross-validation and Bayesian optimization algorithm to obtain the optimal prediction model.

[0011] Step S4: Model interpretability analysis. Using integrated gradient method and attention mechanism weight analysis, key marker sites that significantly contribute to trait prediction are identified, and gene function annotation databases are searched to reveal the potential biological significance of key marker sites.

[0012] Step S5: Multi-trait comprehensive screening. Based on the optimal prediction model, predict the breeding values ​​of multiple traits, use a multi-objective optimization algorithm to calculate the comprehensive selection index of breeding materials, and screen out breeding materials with excellent comprehensive traits.

[0013] Step S6: Screening result verification. Field verification and further evaluation are conducted on the selected breeding materials with excellent comprehensive traits to provide parental materials for the next breeding cycle.

[0014] Further, in step S2, the high-throughput acquisition of the hyperspectral phenotype specifically includes the following steps:

[0015] Step S21: Use a drone equipped with a hyperspectral imaging system to conduct low-altitude remote sensing flight over the experimental field, set the flight altitude, ground resolution, forward overlap rate and lateral overlap rate, and collect hyperspectral images of ginseng at each key growth stage;

[0016] Step S22: Perform radiometric calibration and atmospheric correction on the acquired hyperspectral images to eliminate the influence of illumination conditions and atmospheric environment, obtain surface reflectance data, and generate corrected hyperspectral images;

[0017] Step S23: Based on the boundary vector data of the experimental field, extract the ROI (Region of Interest) corresponding to each breeding material from the corrected hyperspectral image, calculate the average reflectance spectrum of all pixels within the ROI, and obtain the spectral curve of each breeding material;

[0018] Step S24: Preprocess the spectral curves of each breeding material to reduce high-frequency noise, eliminate light scattering effects and baseline drift, and obtain preprocessed spectral data;

[0019] Step S25: Extract commonly used vegetation indices from the preprocessed spectral data, including normalized vegetation index, enhanced vegetation index, and soil-regulated vegetation index, screen characteristic wavelengths, and obtain low-dimensional spectral feature vectors.

[0020] Step S26: Establish a quantitative correlation model for rapid prediction of saponin content in all breeding materials and perform high-throughput augmentation of phenotypic data.

[0021] Furthermore, in step S3, the construction of the multi-trait prediction model specifically includes the following steps:

[0022] Step S31: Input vector construction. Dimensionality reduction is performed on the genotype data to construct genotype feature vectors, which are then fused with low-dimensional spectral feature vectors to construct the input vector, in the following form:

[0023] ;

[0024] In the formula, Represents the input vector. The genotype feature vector represents... Represents a low-dimensional spectral eigenvector;

[0025] Step S32: Construct a multi-trait prediction model based on a multi-task learning framework and deep neural networks, comprising three parts: a shared feature extraction layer, a task-specific layer, and a joint output layer, specifically:

[0026] Step S321: The shared feature extraction layer uses a one-dimensional convolutional layer to extract local features from the input genotype feature vector to obtain a feature map. A multi-head self-attention layer is then added after the one-dimensional convolutional layer to capture long-range dependencies between gene SNP sites and output shared features. The formula used is as follows:

[0027] ;

[0028] ;

[0029] ;

[0030] ;

[0031] ;

[0032] In the formula, This represents the feature map output by a one-dimensional convolutional layer. Represents the linear rectified activation function. This represents a one-dimensional convolution operation. Indicates the convolution kernel weights, Indicates the bias term. This indicates the multi-head self-attention layer. , , These represent the query matrix, key matrix, and value matrix, respectively. , , These represent the corresponding weight matrices. This represents the scaling factor used to stabilize the gradient. Represents the normalized exponential function, This represents the matrix transpose operation;

[0033] Step S322: The task-specific layer consists of fully connected sub-networks, each corresponding to a trait. They share input features and learn trait-specific nonlinear mappings through the fully connected sub-networks. The formula used is as follows:

[0034] ;

[0035] ;

[0036] In the formula, Indicates shared features, Indicates the first The hidden layer output corresponding to the individual traits and These represent the weights and biases of the hidden layer, respectively. and These represent the weights and biases of the output layer, respectively. Indicates the first The first breeding material Predictive values ​​for personality traits;

[0037] Step S323: Construct a joint loss function by comprehensively considering the prediction errors of each trait and the correlation between traits. The formula used is as follows:

[0038] ;

[0039] ;

[0040] ;

[0041] In the formula, Denotes the joint loss function. Represents the regularization term for trait correlation. Represents the coefficient of the regularization term. Indicates the first Mean squared error loss of individual traits Indicates the total number of traits. Indicates the total number of breeding materials. This represents the loss weight for the corresponding trait. and They represent the first and the Pearson correlation coefficient of an individual trait between true and predicted values;

[0042] Step S33: Dataset partitioning, randomly dividing the breeding material population into training set, validation set and test set;

[0043] Step S34: Model training. The Bayesian optimization algorithm is used to automatically search and optimize the hyperparameters of the multi-trait prediction model, select the optimal combination of hyperparameters, and use an early stopping mechanism to prevent overfitting.

[0044] Step S35: Model evaluation. Evaluate the final performance on the test set and calculate the prediction accuracy index for each trait.

[0045] Further, in step S4, the model interpretability analysis specifically includes the following steps:

[0046] Step S41: Calculate the contribution of each input vector to the prediction result using the ensemble gradient method, as shown in the following formula:

[0047] ;

[0048] In the formula, Indicates the first The breeding material was in the first Attribution values ​​for each SNP site, and These represent the original input vector and the baseline input vector, respectively. and Represent the first and second halves of the original input vector and the baseline input vector, respectively. The breeding material was in the first Input values ​​for each SNP site, This represents a multi-trait prediction model. The model outputs the gradient values ​​accumulated with respect to the input through the integration path. Indicates the interpolation path parameters;

[0049] Step S42: Visualize and analyze the weight matrix of the multi-head self-attention layer, and calculate the average attention weight of different SNP sites in each attention head. The formula used is as follows:

[0050] ;

[0051] In the formula, This represents the average attention weight. Indicates the total number of attention heads. Indicates the index of the attention head. Indicates the first The first one in the attention. Weights of each SNP site;

[0052] Step S43: Integrate the gradient attribution value and attention weights to calculate the overall importance score for each SNP site, using the following formula:

[0053] ;

[0054] In the formula, This indicates the overall importance score. This represents the weighted harmonic parameters. Indicates the first The integrated gradient attribution value of each SNP site to the model prediction results;

[0055] Step S44: Sort the importance scores and select the top 1% of SNP sites with the highest scores as key markers;

[0056] Step S45: Reveal the biological interpretability of the model by comparing the selected key marker sites with publicly available gene function annotation databases, labeling the genes they belong to and neighboring genes, and analyzing the potential functions of these genes in ginsenoside biosynthesis and growth and development regulation.

[0057] Furthermore, in step S5, the multi-trait comprehensive screening specifically includes the following steps:

[0058] Step S51: Based on the optimal prediction model, predict the breeding values ​​of each trait for all breeding materials to obtain a prediction value matrix, and set the weight coefficients of each trait according to the breeding objectives, in the following form:

[0059] ;

[0060] In the formula, A matrix representing the predicted values ​​of phenotypic data;

[0061] Step S52: Calculate the comprehensive selection index for each breeding material using the following formula:

[0062] ;

[0063] In the formula, Indicates the first The comprehensive selection index of individual breeding materials Indicates the first Weighting coefficients for individual traits and They represent the first Mean and standard deviation of predicted values ​​for individual traits across all breeding materials;

[0064] Step S53: Combining the trade-offs between traits, a non-dominated sorting genetic algorithm is used for multi-objective optimization, with the predicted values ​​of each trait as the optimization objective, to solve for the optimal material set on the Pareto front.

[0065] Step S54: Combining the comprehensive selection index and Pareto optimization results, select breeding materials with excellent comprehensive performance and reasonable trait configuration, and classify them into grades according to their scores.

[0066] The beneficial effects achieved by the present invention using the above solution are as follows:

[0067] (1) In view of the problems that traditional phenotypic identification methods are inefficient and destructive, and that traditional single-trait models ignore trait associations, this scheme uses UAV hyperspectral remote sensing technology to achieve rapid and non-destructive determination of saponin content, which greatly improves the throughput of phenotypic data acquisition. A multi-task learning architecture and deep neural network are used to construct a multi-trait prediction model. By sharing the representation layer, the genetic association between traits is captured, while retaining the specific information of each trait, so as to achieve collaborative and accurate prediction of multiple traits.

[0068] (2) To address the pain point of poor interpretability of deep learning models, this scheme accurately locates key marker sites by integrating gradient and attention weight analysis, and reveals their biological significance by combining functional annotations, thereby enhancing the transparency and credibility of the model, providing theoretical guidance for molecular marker-assisted breeding, and ensuring that the model has good generalization performance on independent test sets, thus providing a reliable guarantee for practical breeding applications.

[0069] (3) In response to the challenge of comprehensive screening of multiple traits, this scheme combines the prediction results of multiple traits with multi-objective optimization, scientifically balances the trade-offs between various breeding objectives, realizes accurate grading and intelligent screening of breeding materials, significantly improves the scientificity and efficiency of breeding decisions, establishes a complete breeding closed loop, and ensures that the technical scheme can be implemented in actual breeding work and generate continuous technical benefits. Attached Figure Description

[0070] Figure 1 This is a flowchart illustrating the high-throughput intelligent screening method for multi-trait ginseng breeding materials based on deep learning proposed in this invention.

[0071] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0072] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0073] Example 1, see Figure 1 The present invention provides a high-throughput intelligent screening method for multi-trait ginseng breeding materials based on deep learning, the method comprising the following steps:

[0074] Step S1: Breeding data collection, collecting ginseng germplasm resources to construct breeding material populations, performing genotyping and phenotyping on each breeding material, and obtaining genotypic and phenotypic data;

[0075] Step S2: High-throughput acquisition of hyperspectral phenotypes. Hyperspectral images of ginseng in the experimental field are collected using a drone equipped with a hyperspectral imaging system. Low-dimensional spectral feature vectors are obtained through preprocessing and feature extraction. A quantitative correlation model between spectral features and quality traits is established to augment the phenotypic data in a high-throughput manner.

[0076] Step S3: Construction of multi-trait prediction model. A multi-task learning architecture and deep neural network are used to construct a multi-trait prediction model. Based on genotype data and low-dimensional spectral feature vectors, the hyperparameters of the multi-trait prediction model are optimized using ten-fold cross-validation and Bayesian optimization algorithm to obtain the optimal prediction model.

[0077] Step S4: Model interpretability analysis. Using integrated gradient method and attention mechanism weight analysis, key marker sites that significantly contribute to trait prediction are identified, and gene function annotation databases are searched to reveal the potential biological significance of key marker sites.

[0078] Step S5: Multi-trait comprehensive screening. Based on the optimal prediction model, predict the breeding values ​​of multiple traits, use a multi-objective optimization algorithm to calculate the comprehensive selection index of breeding materials, and screen out breeding materials with excellent comprehensive traits.

[0079] Step S6: Screening result verification. Field verification and further evaluation are conducted on the selected breeding materials with excellent comprehensive traits to provide parental materials for the next breeding cycle.

[0080] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the breeding data collection specifically includes the following steps:

[0081] Step S11: Genotype data acquisition and processing, including the following steps:

[0082] Step S111: Collect young ginseng leaf samples and extract genomic DNA using the modified CTAB (hexadecyltrimethylammonium bromide) method;

[0083] Step S112: Sequencing the extracted genomic DNA using whole-genome resequencing technology, with an average sequencing depth of not less than 10×, to obtain raw sequencing data;

[0084] Step S113: Align the raw sequencing data to the ginseng reference genome, and use GATK genome analysis software to perform SNP (single nucleotide polymorphism) detection and genotyping to obtain the raw data of the whole genome SNP marker set;

[0085] Step S114: Perform quality control filtering on the raw data of the whole genome SNP marker set, removing SNP loci with a deletion rate exceeding 20% ​​and a minimum allele frequency below 0.05. Use Beagle software to impute missing genotypes and store the genotype data in matrix form, as follows:

[0086] ;

[0087] In the formula, Representing genotype data in matrix form, Index representing breeding materials, An index representing a SNP site. Indicates the first The breeding material was in the first Genotype encoding of each SNP locus;

[0088] Step S12: Acquisition and processing of tabular data, including the following steps:

[0089] Step S121: Collect dynamic data on the key growth stages of ginseng over time, including field phenotypic data for the flowering period, green fruit stage, red fruit stage, and harvest period;

[0090] Step S122: Measure the appearance morphology data of ginseng, including plant height, stem diameter, leaf area, and chlorophyll content. Each breeding material is measured three times, and the average value is taken as the phenotypic value.

[0091] Step S123: Collect ginseng yield composition data. During the harvest period, collect ginseng taproots, wash and dry them, weigh the fresh weight and dry weight of a single root, and measure the length and diameter of the taproot.

[0092] Step S124: Collect ginseng quality data, take dried ginseng root powder, and use high performance liquid chromatography to determine the content of 9 major ginsenosides, including Rg1, Re, Rf, Rg2, Ro, Rc, Rb2, Rb3, and Rd, and calculate the total saponin content.

[0093] Step S125: Perform outlier detection and normality testing on the time-dimensional dynamic data, appearance morphology data, yield composition data, and quality data. Perform data transformation and standardization on phenotypic data that do not conform to a normal distribution, and store the phenotypic data in matrix form as follows:

[0094] ;

[0095] In the formula, Representing tabular data in matrix form, Represents a trait index. Indicates the first The breeding material was in the first Standardized phenotypic values ​​of individual traits.

[0096] Example 3, see Figure 1 This embodiment is based on the above embodiment. In step S2, the high-throughput acquisition of the hyperspectral phenotype specifically includes the following steps:

[0097] Step S21: Use a drone equipped with a hyperspectral imaging system to conduct low-altitude remote sensing flight over the experimental field. The flight altitude is set to 30 meters, the ground resolution reaches 5 centimeters, and the forward overlap rate and lateral overlap rate are both not less than 80%. Collect hyperspectral images of ginseng at each key growth stage.

[0098] Step S22: Perform radiometric calibration and atmospheric correction on the acquired hyperspectral images to eliminate the influence of illumination conditions and atmospheric environment, obtain surface reflectance data, and generate corrected hyperspectral images;

[0099] Step S23: Based on the boundary vector data of the experimental field, extract the ROI corresponding to each breeding material from the corrected hyperspectral image, calculate the average reflectance spectrum of all pixels in the ROI, and obtain the spectral curve of each breeding material with a wavelength range of 400-1000nm.

[0100] Step S24: Preprocess the spectral curves of each breeding material, use Savitzky-Golay filtering to reduce high-frequency noise, set the window width to 9 and the polynomial order to 2; use multivariate scattering correction to eliminate light scattering effects; use standard normal variable transformation to eliminate baseline drift, and obtain the preprocessed spectral data;

[0101] Step S25: Extract commonly used vegetation indices from the preprocessed spectral data, including normalized vegetation index, enhanced vegetation index, and soil-regulated vegetation index. Use the continuous projection algorithm to screen characteristic wavelengths from the full-band spectrum to obtain low-dimensional spectral feature vectors.

[0102] Step S26: Establish a quantitative correlation model. Use partial least squares regression and Gaussian process regression to construct prediction models respectively. Input low-dimensional spectral feature vectors and output the content of 9 major ginsenosides and total saponin content. Evaluate the accuracy of the prediction model through 10-fold cross-validation. Select the optimal prediction model as the quantitative correlation model for rapid prediction of saponin content in all breeding materials and perform high-throughput augmentation of phenotypic data.

[0103] Example 4, see Figure 1 This embodiment is based on the above embodiment. In step S3, the construction of the multi-trait prediction model specifically includes the following steps:

[0104] Step S31: Input vector construction. Dimensionality reduction is performed on the genotype data to construct genotype feature vectors, which are then fused with low-dimensional spectral feature vectors to construct the input vector, in the following form:

[0105] ;

[0106] In the formula, Represents the input vector. The genotype feature vector represents... Represents a low-dimensional spectral eigenvector;

[0107] Step S32: Construct a multi-trait prediction model based on a multi-task learning framework and deep neural networks, comprising three parts: a shared feature extraction layer, a task-specific layer, and a joint output layer, specifically:

[0108] Step S321: The shared feature extraction layer uses a one-dimensional convolutional layer to extract local features from the input genotype feature vector to obtain a feature map. A multi-head self-attention layer is then added after the one-dimensional convolutional layer to capture long-range dependencies between gene SNP sites and output shared features. The formula used is as follows:

[0109] ;

[0110] ;

[0111] ;

[0112] ;

[0113] ;

[0114] In the formula, This represents the feature map output by a one-dimensional convolutional layer. Represents the linear rectified activation function. This represents a one-dimensional convolution operation. Indicates the convolution kernel weights, Indicates the bias term. This indicates the multi-head self-attention layer. , , These represent the query matrix, key matrix, and value matrix, respectively. , , These represent the corresponding weight matrices. This represents the scaling factor used to stabilize the gradient. Represents the normalized exponential function, This represents the matrix transpose operation;

[0115] Step S322: The task-specific layer consists of fully connected sub-networks, each corresponding to a trait. They share input features and learn trait-specific nonlinear mappings through the fully connected sub-networks. The formula used is as follows:

[0116] ;

[0117] ;

[0118] In the formula, Indicates shared features, Indicates the first The hidden layer output corresponding to the individual traits and These represent the weights and biases of the hidden layer, respectively. and These represent the weights and biases of the output layer, respectively. Indicates the first The first breeding material Predictive values ​​for personality traits;

[0119] Step S323: Construct a joint loss function by comprehensively considering the prediction errors of each trait and the correlation between traits. The formula used is as follows:

[0120] ;

[0121] ;

[0122] ;

[0123] In the formula, Denotes the joint loss function. Represents the regularization term for trait correlation. Represents the coefficient of the regularization term. Indicates the first Mean squared error loss of individual traits Indicates the total number of traits. Indicates the total number of breeding materials. This represents the loss weight for the corresponding trait. and They represent the first and the Pearson correlation coefficient of an individual trait between true and predicted values;

[0124] Step S33: Dataset partitioning. The breeding material population is randomly divided into training set, validation set and test set in a ratio of 7:2:1.

[0125] Step S34: Model training, including the following steps:

[0126] Step S341: Divide the training set into 10 parts, use 9 parts for training and 1 part for validation in turn, and use 10-fold cross-validation to calculate the average prediction accuracy.

[0127] Step S342: The Bayesian optimization algorithm is used to automatically search and optimize the hyperparameters of the multi-trait prediction model. The hyperparameters include the number of convolutional kernels, the number of attention heads, the dimension of hidden layers, the learning rate, the dropout ratio, and the regularization coefficient. The average prediction accuracy on the validation set is used as the objective function. The optimization is performed 50 times to select the optimal combination of hyperparameters.

[0128] Step S343: Use an early stopping mechanism to prevent overfitting. Set the training to stop when the validation set loss no longer decreases after 10 consecutive epochs and restore to the optimal hyperparameter combination.

[0129] Step S35: Model evaluation. Evaluate the final performance on the test set and calculate the prediction accuracy metrics for each trait, including Pearson correlation coefficient, root mean square error, and mean absolute error.

[0130] By performing the aforementioned operations, this solution addresses the problems of low efficiency, high destructiveness, and neglect of trait associations in traditional phenotypic identification methods and traditional single-trait models. It utilizes UAV hyperspectral remote sensing technology to achieve rapid and non-destructive determination of saponin content, significantly increasing the throughput of phenotypic data acquisition. A multi-task learning architecture and deep neural networks are used to construct a multi-trait prediction model. By sharing the representation layer, the genetic associations between traits are captured, while retaining the specific information of each trait, thus achieving collaborative and accurate prediction of multiple traits.

[0131] Example 5, see Figure 1 This embodiment is based on the above embodiment. In step S4, the model interpretability analysis specifically includes the following steps:

[0132] Step S41: Calculate the contribution of each input vector to the prediction result using the ensemble gradient method, as shown in the following formula:

[0133] ;

[0134] In the formula, Indicates the first The breeding material was in the first Attribution values ​​for each SNP site, and These represent the original input vector and the baseline input vector, respectively. and Represent the first and second halves of the original input vector and the baseline input vector, respectively. The breeding material was in the first Input values ​​for each SNP site, This represents a multi-trait prediction model. The model outputs the gradient values ​​accumulated with respect to the input through the integration path. Indicates the interpolation path parameters;

[0135] Step S42: Visualize and analyze the weight matrix of the multi-head self-attention layer, and calculate the average attention weight of different SNP sites in each attention head. The formula used is as follows:

[0136] ;

[0137] In the formula, This represents the average attention weight. Indicates the total number of attention heads. Indicates the index of the attention head. Indicates the first The first one in the attention. Weights of each SNP site;

[0138] Step S43: Integrate the gradient attribution value and attention weights to calculate the overall importance score for each SNP site, using the following formula:

[0139] ;

[0140] In the formula, This indicates the overall importance score. This represents the weighted harmonic parameters. Indicates the first The integrated gradient attribution value of each SNP site to the model prediction results;

[0141] Step S44: Sort the importance scores and select the top 1% of SNP sites with the highest scores as key markers;

[0142] Step S45: Reveal the biological interpretability of the model by comparing the selected key marker sites with publicly available gene function annotation databases, labeling the genes they belong to and neighboring genes, and analyzing the potential functions of these genes in ginsenoside biosynthesis and growth and development regulation.

[0143] By performing the aforementioned operations, this solution addresses the pain point of poor interpretability in deep learning models. It accurately locates key marker sites through integrated gradient and attention weight analysis, and reveals their biological significance by combining functional annotations. This enhances the transparency and credibility of the model, provides theoretical guidance for molecular marker-assisted breeding, and ensures that the model has good generalization performance on independent test sets, thus providing a reliable guarantee for practical breeding applications.

[0144] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S5, the multi-trait comprehensive screening specifically includes the following steps:

[0145] Step S51: Based on the optimal prediction model, predict the breeding values ​​of each trait for all breeding materials to obtain a prediction value matrix, and set the weight coefficients of each trait according to the breeding objectives, in the following form:

[0146] ;

[0147] In the formula, A matrix representing the predicted values ​​of phenotypic data;

[0148] Step S52: Calculate the comprehensive selection index for each breeding material using the following formula:

[0149] ;

[0150] In the formula, Indicates the first The comprehensive selection index of individual breeding materials Indicates the first Weighting coefficients for individual traits and They represent the first Mean and standard deviation of predicted values ​​for individual traits across all breeding materials;

[0151] Step S53: Combining the trade-offs between traits, a non-dominated sorting genetic algorithm is used for multi-objective optimization, with the predicted values ​​of each trait as the optimization objective, to solve for the optimal material set on the Pareto front.

[0152] Step S54: Combining the comprehensive selection index and Pareto optimization results, select breeding materials with excellent overall performance and reasonable trait configuration, and classify them into three levels: excellent, good, and candidate according to their scores.

[0153] Example 7, see Figure 1 This embodiment is based on the above embodiment. In step S6, the verification of the screening results specifically includes the following steps:

[0154] Step S61: For the selected superior and excellent breeding materials, field verification trials will be conducted in the next growing season. A randomized block design will be used with three replicates to comprehensively measure the phenotypic values ​​of each trait.

[0155] Step S62: Compare and analyze the measured values ​​in the field with the predicted values ​​of the model, calculate and verify the accuracy, and evaluate the generalization ability of the model in independent years and environments;

[0156] Step S63: For the verified superior materials, accelerate the propagation of seeds and seedlings, and enter the variety comparison test and regional trial;

[0157] Step S64: Use the selected superior parental materials for the next round of hybridization to build a new breeding population, accumulate superior alleles, and continuously improve breeding efficiency.

[0158] By performing the aforementioned operations, this solution addresses the challenge of comprehensive screening of multiple traits. It combines multi-trait prediction results with multi-objective optimization, scientifically balances the trade-offs between various breeding objectives, achieves precise grading and intelligent screening of breeding materials, significantly improves the scientificity and efficiency of breeding decisions, establishes a complete breeding closed loop, and ensures that the technical solution can be applied in actual breeding work, generating continuous technical benefits.

[0159] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0160] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0161] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A high-throughput intelligent screening method for multi-trait ginseng breeding materials based on deep learning, characterized by: The method includes the following steps: Step S1: Breeding data collection, collecting ginseng germplasm resources to construct breeding material populations, performing genotyping and phenotyping on each breeding material, and obtaining genotypic and phenotypic data; Step S2: High-throughput acquisition of hyperspectral phenotypes. Hyperspectral images of ginseng in the experimental field are collected using a drone equipped with a hyperspectral imaging system. Low-dimensional spectral feature vectors are obtained through preprocessing and feature extraction. A quantitative correlation model between spectral features and quality traits is established to augment the phenotypic data in a high-throughput manner. Step S3: Construction of multi-trait prediction model. A multi-task learning architecture and deep neural network are used to construct a multi-trait prediction model. Based on genotype data and low-dimensional spectral feature vectors, the hyperparameters of the multi-trait prediction model are optimized using ten-fold cross-validation and Bayesian optimization algorithm to obtain the optimal prediction model. Step S4: Model interpretability analysis. Using integrated gradient method and attention mechanism weight analysis, key marker sites that significantly contribute to trait prediction are identified, and gene function annotation databases are searched to reveal the potential biological significance of key marker sites. Step S5: Multi-trait comprehensive screening. Based on the optimal prediction model, predict the breeding values ​​of multiple traits, use a multi-objective optimization algorithm to calculate the comprehensive selection index of breeding materials, and screen out breeding materials with excellent comprehensive traits. Step S6: Screening result verification. Field verification and further evaluation are conducted on the selected breeding materials with excellent comprehensive traits to provide parental materials for the next breeding cycle.

2. The high-throughput intelligent screening method for multi-trait ginseng breeding materials based on deep learning according to claim 1, characterized in that: In step S2, the high-throughput acquisition of the hyperspectral phenotype specifically includes the following steps: Step S21: Use a drone equipped with a hyperspectral imaging system to conduct low-altitude remote sensing flight over the experimental field and collect hyperspectral images of ginseng at each key growth stage; Step S22: Perform radiometric calibration and atmospheric correction on the acquired hyperspectral images to eliminate the influence of illumination conditions and atmospheric environment, obtain surface reflectance data, and generate corrected hyperspectral images; Step S23: Based on the boundary vector data of the experimental field, extract the ROI corresponding to each breeding material from the corrected hyperspectral image, calculate the average reflectance spectrum of all pixels within the ROI, and obtain the spectral curve of each breeding material; Step S24: Preprocess the spectral curves of each breeding material to reduce high-frequency noise, eliminate light scattering effects and baseline drift, and obtain preprocessed spectral data; Step S25: Extract commonly used vegetation indices from the preprocessed spectral data, screen characteristic wavelengths, and obtain low-dimensional spectral feature vectors; Step S26: Establish a quantitative correlation model for rapid prediction of saponin content in all breeding materials and perform high-throughput augmentation of phenotypic data.

3. The high-throughput intelligent screening method for multi-trait ginseng breeding materials based on deep learning according to claim 1, characterized in that: In step S3, the construction of the multi-trait prediction model specifically includes the following steps: Step S31: Input vector construction. Dimensionality of genotype data is reduced to construct genotype feature vectors, which are then fused with low-dimensional spectral feature vectors to construct the input vector. Step S32: Construct a multi-trait prediction model based on a multi-task learning framework and deep neural networks, comprising three parts: a shared feature extraction layer, a task-specific layer, and a joint output layer, specifically: Step S321: The shared feature extraction layer uses a one-dimensional convolutional layer to extract local features from the input genotype feature vector to obtain a feature map. A multi-head self-attention layer is added after the one-dimensional convolutional layer to capture the long-range dependencies between gene SNP sites and output the shared features. Step S322: The task-specific layer consists of fully connected sub-networks, each sub-network corresponding to a trait. The inputs share features, and the fully connected sub-networks learn trait-specific nonlinear mappings. Step S323: Construct a joint loss function by comprehensively considering the prediction errors of each trait and the correlation between traits; Step S33: Dataset partitioning, randomly dividing the breeding material population into training set, validation set and test set; Step S34: Model training. The Bayesian optimization algorithm is used to automatically search and optimize the hyperparameters of the multi-trait prediction model, select the optimal combination of hyperparameters, and use an early stopping mechanism to prevent overfitting. Step S35: Model evaluation. Evaluate the final performance on the test set and calculate the prediction accuracy index for each trait.

4. The high-throughput intelligent screening method for multi-trait ginseng breeding materials based on deep learning according to claim 1, characterized in that: In step S4, the model interpretability analysis specifically includes the following steps: Step S41: Calculate the contribution of each input vector to the prediction result using the ensemble gradient method; Step S42: Visualize and analyze the weight matrix of the multi-head self-attention layer, and calculate the average attention weight of different SNP sites in each attention head; Step S43: Integrate gradient attribution values ​​and attention weights to calculate the overall importance score for each SNP site; Step S44: Sort the importance scores and select the top 1% of SNP sites with the highest scores as key markers; Step S45: Reveal the biological interpretability of the model by comparing the selected key marker sites with publicly available gene function annotation databases, labeling the genes they belong to and neighboring genes, and analyzing the potential functions of these genes in ginsenoside biosynthesis and growth and development regulation.

5. The high-throughput intelligent screening method for multi-trait ginseng breeding materials based on deep learning according to claim 1, characterized in that: In step S5, the multi-trait comprehensive screening specifically includes the following steps: Step S51: Based on the optimal prediction model, predict the breeding values ​​of each trait of all breeding materials to obtain the prediction value matrix, and set the weight coefficients of each trait according to the breeding objectives. Step S52: Calculate the overall selection index for each breeding material; Step S53: Combining the trade-offs between traits, a non-dominated sorting genetic algorithm is used for multi-objective optimization, with the predicted values ​​of each trait as the optimization objective, to solve for the optimal material set on the Pareto front. Step S54: Combining the comprehensive selection index and Pareto optimization results, select breeding materials with excellent comprehensive performance and reasonable trait configuration, and classify them into grades according to their scores.

Images