High-quality and high-yield alfalfa hybrid combination screening method fusing multiple technologies

Abstract

The invention relates to the technical field of alfalfa screening, discloses a high-quality and high-yield alfalfa hybrid combination screening method fusing multiple technologies, provides a multi-dimensional coding and double-elite evolution mechanism, constructs an improved non-dominated sorting genetic algorithm-II, and provides a high-quality and high-yield alfalfa hybrid combination screening method by coding a hybrid combination into a continuous real number vector. A chromosome expression mode decoupled from a problem structure is constructed, a double-elite mechanism of elite retention and offset mating is introduced, the retention and heredity ability of a high-quality solution is improved, offset crossover and jitter operation is used, premature convergence is avoided, efficient global search of a strain combination space is achieved, and a high-quality solution is obtained. Meanwhile, a multi-omics association model of genes, metabolites and phenotypes is provided, transcriptome, metabolome and molecular marker data are comprehensively utilized, a multi-dimensional character prediction model is constructed, early-stage accurate screening of filial generations is achieved through correlation analysis and marker verification, the error selection rate is reduced, and the breeding period is shortened.

Description

Technical Field

[0001] This invention relates to the field of alfalfa screening technology, and in particular to a screening method for high-quality, high-yield alfalfa hybrid combinations that integrates multiple technologies. Background Technology

[0002] Alfalfa, as one of the world's most important legume forage crops, plays a crucial role in the development of animal husbandry due to its high quality and yield. Currently, the breeding of high-yielding and high-quality alfalfa varieties mainly adopts a combination of traditional hybridization breeding and systematic selection. However, this approach suffers from problems such as low efficiency in combination screening, long breeding cycles, and poor accuracy in trait prediction. In traditional hybridization breeding, parental selection relies on experience, and hybrid combinations require full combination trials, resulting in a large number of combinations with a low success rate. Furthermore, there is a lack of effective process simulation and early screening mechanisms. At the same time, traits such as alfalfa fresh yield and crude protein content are significantly affected by the culture environment, and traditional methods cannot dynamically predict the regulatory effects of the environment on traits, leading to insufficient accuracy in screening superior strains. Although big data and multi-omics technologies have been applied in crop breeding... However, limitations still exist in the field of alfalfa breeding: 1. Existing genetic algorithms are difficult to adapt to the high-dimensional optimization requirements of alfalfa with multiple traits, and are prone to premature convergence; the integration of multi-omics data is insufficient, and it is impossible to construct accurate gene-phenotype associations. 2. The lack of an environment-trait coupling model makes it impossible to achieve intelligent optimization of culture conditions. Summary of the Invention

[0003] The technical problem to be solved by this invention is that the existing technology has the disadvantage that genetic algorithms are difficult to adapt to the high-dimensional optimization requirements of alfalfa with multiple traits and are prone to premature convergence. To this end, we propose a screening method for high-quality and high-yield alfalfa hybrid combinations that integrates multiple technologies.

[0004] To achieve the above objectives, this application adopts the following technical solution: a method for screening high-quality, high-yield alfalfa hybrid combinations integrating multiple technologies, comprising: Step S1: Collect basic data of alfalfa strains, hybridization process data, culture environment parameters and external reference data, and perform standardization and structuring processing using standardized score normalization and one-bit effective coding; Step S2: The improved non-dominated sorting genetic algorithm-II is used to optimize the hybridization combination through multi-objective encoding. The fitness function is constructed with fresh grass yield, crude protein content and salt tolerance as the core objectives. The optimal hybridization scheme is generated by combining the dual elite retention mechanism and the adaptive crossover mutation strategy. Step S3: Construct a gene-metabolite-phenotype association model through transcriptome differential analysis, metabolome key substance detection, and simple repetitive sequence molecular marker verification to achieve early screening of hybrid offspring; Step S4: Construct an environmental dynamic model based on a long short-term memory neural network to simulate the growth performance of strains under soil moisture content, salinity, and temperature gradients, and search for the optimal environmental regulation strategy to maximize comprehensive traits. Step S5: Construct a multi-index comprehensive scoring system based on the analytic hierarchy process (AHP), allocate fresh grass yield, crude protein content, salt tolerance, and reproductive weight, identify abnormal samples and backtrack to adjust the parameters of the preceding model, and output high-performance hybridization combinations.

[0005] In step S1: Basic data on the strains include the genetic background, agronomic traits, and stress resistance data of the parent alfalfa varieties Zhongmu 3 and WL323HQ; Data on the hybridization process includes the fruit set rate of individual hybrids, Generation coefficient, Separation ratio; The cultivation environment parameters include drought stress gradient, salinity gradient, and temperature; External reference data include alfalfa, expression patterns of stress-resistance genes, and multi-trait breeding cases in legumes.

[0006] In step S2: The hybridization combination is encoded as a chromosome vector X represented by 12-dimensional continuous real numbers; The chromosome vector X is decoded into a specific combination strategy and fitness score by an encoder. The objective function is constructed with fresh grass yield over-parent rate, crude protein content and salt tolerance as the evaluation core. By introducing parental gene expression data, simulating gene recombination behavior during hybridization of different strains, predicting the phenotypic probability distribution of offspring, and forming a dynamic hybridization simulation model; Employing a dual-elite evolutionary mechanism, each generation retains the previous generation's fitness. A select few individuals, and with probability >0.5 Perform biased crossover: Offspring genes are distributed with probability. Inheriting the elite parent's gene 1 with a probability of 1- Inherited two genes from non-elite parents; As the number of iterations progresses, parameters such as the elite ratio and mutation probability are dynamically adjusted to adapt to the search progress.

[0007] The objective function in step S2 is: ; in, This represents the fitness score corresponding to chromosome X. The percentage of offspring fresh grass yield exceeding the parental average is the predicted rate. Standardized values ​​of crude protein content in the table. Represents salt tolerance index, , , These represent the weight parameters of the objective function, which are set according to task requirements. The improved non-dominated sorting genetic algorithm-II also includes: Jittering mechanism: applying probability to elite individuals Apply a disturbance; Reset mechanism: When there is no improvement in fitness after a series of preset generations, remove non-elite individuals and regenerate random chromosomes.

[0008] Step S3 involves constructing a correlation model through combined analysis of transcriptomics, metabolomics, and molecular markers, including: We screened differentially expressed genes related to yield and quality by transcriptome sequencing, detected the content of key metabolites by metabolomics, and constructed a hybrid dataset by combining SSR molecular marker data. Correlation analysis was used to fit the gene-metabolite-phenotype association relationship, resulting in a gene-metabolite-phenotype association model. Threshold screening and marker validation were used to form an early screening mechanism, which efficiently simulated the phenotypic performance of hybrid offspring in the model.

[0009] Step S3 also includes: Core genes are screened from differentially expressed gene sets to serve as initial targets for early screening. Within the association model, a threshold method is used to screen hybrid offspring that meet the target trait, and the screening results are optimized through molecular marker validation. The generated compliant lines are further fed back into the hybridization process to produce new data. This forms a closed data loop.

[0010] In step S4: Long Short-Term Memory Neural Network: The input to the environmental dynamics model is soil moisture content, sodium chloride concentration, and temperature, and the output is the predicted values ​​of fresh grass yield and crude protein content. Environmental dynamics model: trained using field data from multiple regions, enabling accurate prediction of strain growth performance under different environments; The optimal environmental control strategy is to regulate the changes in soil moisture content, salinity, and temperature at different stages of the seedling stage.

[0011] In step S5: The formula for the comprehensive scoring system is: ; in This represents the actual yield of fresh grass. This represents the minimum yield of fresh grass. This represents the maximum yield of fresh grass. This represents the maximum crude protein content. This represents the minimum crude protein content. This is the measured value of crude protein content. This is the measured value of salt tolerance. This is the optimal value for the reproductive period. These are measured values ​​during the reproductive period; Identify abnormal samples: Trigger a backtracking adjustment mechanism for abnormal samples to adjust the parameters of the previous screening process to optimize the subsequent screening results.

[0012] In step S5: Each score is standardized. When the overall score of the strain is lower than the preset threshold or the core traits are not met, the system automatically backtracks to the parent selection or hybridization design stage and updates the parameters of the genetic algorithm or association model to optimize the prediction quality of hybridization combinations.

[0013] A screening system for high-quality, high-yield alfalfa hybrid combinations integrating multiple technologies is used to implement a screening method for high-quality, high-yield alfalfa hybrid combinations integrating multiple technologies, including: Data processing module: used for the acquisition, standardization, and structured processing of multi-source data; Hybridization optimization module: Runs an improved multi-objective genetic algorithm to generate the optimal hybridization scheme; Multi-omics screening module: Integrates multi-omics and molecular marker data to achieve early screening; Environmental simulation module: Constructs dynamic models based on recurrent neural networks and outputs environmental control strategies; Performance evaluation module: Outputs high-performance hybridization combinations through hierarchical evaluation and backtracking adjustment. The technical effects and advantages of this invention are as follows: This invention proposes a multi-dimensional encoding and dual-elite evolution mechanism, constructs an improved non-dominated sorting genetic algorithm-II, encodes hybridization combinations as continuous real vectors, constructs a chromosome expression mode decoupled from the problem structure, introduces a dual-elite mechanism of elite retention and biased mating to improve the retention and genetic ability of high-quality solutions, and uses biased crossover and jitter operations to avoid premature convergence and achieve efficient global search of the strain combination space. This invention proposes a multi-omics association model of genes, metabolites, and phenotypes. It comprehensively utilizes transcriptomic, metabolomic, and molecular marker data to construct a multi-dimensional trait prediction model. Through correlation analysis and marker validation, it enables early and accurate screening of hybrid offspring, reduces the false selection rate, and shortens the breeding cycle. This invention proposes a multi-index evaluation and backtracking mechanism based on the analytic hierarchy process (AHP), constructing a multi-dimensional comprehensive scoring system and incorporating penalty term design to automatically identify low-performance strains. By backtracking and adjusting parameters in the upstream optimization process, the breeding strategy is iteratively improved, thereby increasing the acquisition rate of superior strains. Attached Figure Description

[0014] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts: Figure 1 This is a schematic diagram of a screening method for high-quality, high-yield alfalfa hybrid combinations that integrates multiple technologies, according to the present invention. Figure 2 This is a schematic diagram of a screening system for high-quality, high-yield alfalfa hybrid combinations that integrates multiple technologies, according to the present invention. Detailed Implementation

[0015] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention. Example

[0016] The high-yield, high-quality alfalfa strain JAUA086 was developed by collecting seeds from high-yield cultivated alfalfa strain B218 and high-crude-protein wild alfalfa strain C356, separating individual plants, obtaining homozygous strains, and hybridizing them to create true hybrid strains. The hybrid strain D48, which had the highest fresh grass yield and crude protein content exceeding that of its parent strain, was selected. The homozygous strain obtained from the collected seeds was then backcrossed with the homozygous strain of parent strain C356, resulting in the hybrid strain JAUA086, which had both higher fresh grass yield and crude protein content than its parent strain. This hybrid strain JAUA086 was cultured in a substrate containing 50% corn stalks, 30% soybean meal, 15% wheat bran, 5% inorganic salt mixture, 2.0 g / kg ferrous sulfate, and 60% moisture content.

[0017] The high-yield alfalfa strain of this invention is created by starting from its own specific traits and improving the variety through genetic breeding and matching cultivation methods. In contrast, traditional alfalfa breeding relies more on empirical selection and does not build trait associations at the gene level, resulting in limitations such as low screening efficiency and unstable traits.

[0018] Based on the evaluation of a large number of alfalfa germplasm resources, this invention selects high-yielding and high-crude-protein strains as hybrid parents, making it easier to create superior high-yielding strains. Combined with a suitable culture medium formula, it can cultivate strains with stable target traits. Compared with hybridization breeding technology and systematic selection technology alone, this invention is more targeted and efficient, improving breeding efficiency. This invention not only improves the genetic traits of alfalfa, but also enhances the high-yielding and high-quality traits by changing the culture conditions, significantly improving the accuracy of strain selection and providing a new approach for the breeding of functional strains of forage grass.

[0019] This invention utilizes the integration of multiple technologies to construct a high-quality, high-yield alfalfa hybrid combination screening system. From data collection, combination optimization, early screening to environmental simulation and performance evaluation, a closed-loop automatic optimization process is formed, providing an intelligent decision-making tool for alfalfa and other forage breeding. It significantly improves efficiency and breeding accuracy in hybrid combination optimization, early screening, and environmental control.

[0020] Reference Figure 1 As shown, this invention provides a technical solution: a method for screening high-quality, high-yield alfalfa hybrid combinations that integrates multiple technologies, comprising: Step S1: Collect basic data of the strain, hybridization process data, culture environment parameters and external reference data, and perform standardization and structuring processing; Step S1 is used for multi-source data acquisition and preprocessing. To build an intelligent screening system for high-quality, high-yield alfalfa hybrid combinations, a complete, standardized, and structured data foundation is first required. Through systematic acquisition and cleaning of multi-source heterogeneous data, high-quality support is provided for subsequent hybridization optimization, model prediction, and performance evaluation. This mainly includes the following steps: Step S1.1: Collection of four types of data Step S1.1.1: Basic data collection of strains Genetic background and agronomic traits of key parental strains are collected as basic inputs for hybridization design.

[0021] Genetic information, such as genotype data of parents B218 and C356, SSR marker sites (Xmsat106, Xmsat448), gene expression data, etc., are used to help infer the genetic expression ability of hybrid offspring.

[0022] Agronomic trait data include fresh grass yield, crude protein content, salt tolerance, and growth period. For example, the fresh grass yield and crude protein content of high-yielding alfalfa cultivated strain B218; and the fresh grass yield and crude protein content of high-crude-protein wild alfalfa strain C356, which serve as reference baselines for offspring over-parent performance.

[0023] Step S1.1.2: Data Acquisition During Hybridization Data was collected for key steps in each round of hybridization experiments, including operation records and experimental result indicators.

[0024] Single-strain isolation data: Record the number of homozygous strains obtained from the parent population in each round of experiments (e.g., 18 strains obtained in the first round) to provide a diversity basis for subsequent combinatorial design.

[0025] Hybridization combination design: Record the number of combinations generated in each round of hybridization (e.g., 81 combinations are generated in the first round), and combine them with naming rules to make them easy to track and manage.

[0026] Hybrid authenticity verification data: Collect purity detection data of hybrid offspring between different strain combinations (such as SSR marker verification results) to eliminate non-authentic hybrid combinations.

[0027] Step S1.1.3: Collection of culture environment and condition parameters We collected key environmental factors that influence strain growth and trait expression to support subsequent environmental modeling and simulation.

[0028] Culture medium formulation composition: Record the proportions of basic materials (such as corn stalks, soybean meal, wheat bran, inorganic salt mixture, etc.) as input parameters for the culture environment.

[0029] Iron ion concentration setting: This includes the set concentration gradient, which is used as an input variable to associate with the target trait in the prediction model.

[0030] Temperature and humidity conditions: including continuous or phased environmental control data such as seedling temperature and flowering humidity.

[0031] Step S1.1.4: External Reference Data Acquisition Technical routes and successful cases of high-yield and high-quality breeding of other forage grasses (such as alfalfa and sweet clover) are introduced as prior information for system design and model training.

[0032] Functional breeding methods include: high-yield regulation, selection processes to enhance protein synthesis, and metabolic pathway analysis strategies.

[0033] Breeding evaluation index system: Refer to the evaluation methods used in multi-trait evaluation, compare and absorb its effective experience.

[0034] Step S1.2: Data Preprocessing Flow To ensure data availability and model training effectiveness, the collected multi-source data undergoes standardization and structuring processing, including: Missing value imputation and outlier removal (e.g., fresh grass yield below the minimum value is directly removed); Numerical normalization processing (applicable to various continuous indicators, such as yield, concentration, etc.); Category-based feature coding (such as strain number, hybridization combination type); Time series alignment and stage labeling (for reproductive period and dynamic simulation modeling). Through the above multidimensional data acquisition and preprocessing operations, the system has constructed a high-quality data foundation for strain breeding that integrates genetics, physiology, environment and process behavior, laying a solid digital platform foundation for subsequent algorithm optimization and performance improvement. An improved non-dominated sorting genetic algorithm-II is used to encode and optimize the hybridization combination, and the optimal hybridization scheme is generated by combining a dual elite evolution mechanism and a dynamic parameter adjustment strategy. To effectively improve the efficiency of hybridization breeding of high-yield alfalfa strains and reduce ineffective combinations and experimental costs, this system, based on data collection, further introduces an improved non-dominated sorting genetic algorithm-II to construct an intelligent hybridization optimization mechanism and process simulation model. By comprehensively considering three objectives—"fresh grass yield over-parent rate," "crude protein content," and "salt tolerance"—the system dynamically optimizes hybridization combination strategies and backcross generations, significantly improving the accuracy of screening true hybrid offspring.

[0035] Step S2 is used to optimize the hybridization combination and simulate the process. The implementation of the improved non-dominated sorting genetic algorithm-II includes: The hybridization combination is encoded as a chromosome vector X represented by 12-dimensional continuous real numbers; The chromosome vector X is decoded into a specific combination strategy and fitness score by an encoder. The objective function is constructed with fresh grass yield over-parent rate, crude protein content and salt tolerance as the evaluation core. By introducing parental gene expression data, simulating gene recombination behavior during hybridization of different strains, predicting the phenotypic probability distribution of offspring, and forming a dynamic hybridization simulation model; Employing a dual-elite evolutionary mechanism, each generation retains the previous generation's fitness. A select few individuals, and with probability >0.5 Perform biased crossover operation, and offspring genes are distributed with probability. Inheriting the elite parent's gene 1 with a probability of 1- Inherited two genes from non-elite parents; As the iterations progress, parameters such as the elite ratio and mutation probability are dynamically adjusted to adapt to the search progress. Step S2 specifically includes the following steps: Step S2.1: Hybrid Combination Optimization Strategy Traditional hybridization methods involve a large number of combinations (e.g., 81 combinations in the first round), but the actual hybridization success rate is low (e.g., only 36 true hybrid strains were screened). Therefore, this system adopts the following strategy: Authenticity verification and screening: Using SSR markers to verify the results, combinations that meet the purity standards are prioritized for retention, while non-authentic hybridization combinations are initially excluded.

[0036] Dynamic adjustment of backcross strategy: Based on the mid-term test results, dynamically increase the number of backcrosses of offspring with obvious target traits, such as "hybrid strain D48 and high crude protein parent C356", to ensure the genetic stability of the target trait.

[0037] Multi-objective optimization: By constructing a joint objective function, the stable performance of agronomic traits such as fresh grass yield, crude protein content, and salt tolerance is balanced.

[0038] Step S2.2: Constructing an improved non-dominated sorting genetic algorithm-II The system employs an improved non-dominated sorting genetic algorithm-II for combinatorial optimization. Its core feature is a multi-dimensional chromosome combination dual-elite evolution mechanism, which has good search space coverage and global optimization efficiency.

[0039] Chromosome vectors are virtual codes designed in the Improved Non-Dominated Sorting Genetic Algorithm-II to represent the optimization strategy of crossover combinations. They are a mathematical abstract model, with each dimension... For each breeding decision variable, a decoder can transform it into an actionable hybridization breeding strategy (e.g., parental selection, number of backcrosses, recommended environmental parameters such as iron concentration in the culture medium), ultimately outputting the optimal strategy to guide experimental design. A chromosome vector is defined as a 12-dimensional real-number vector representing each individual. ; X: Chromosome vector, used to encode the selection order and strategy parameters of strain combinations. 12 is the encoding dimension, which is determined by the number of decision variables to be optimized (such as parental selection, number of backcrosses, environmental parameter weights, etc.). Let be the i-th encoded value, representing the priority or proportional allocation parameter in the decoder. It is a continuous real number in the interval (0,1], representing the probability weight or priority of a certain decision. The decoder decodes the chromosome vector X into a specific combination of policy and fitness score, for example, through threshold decoding (e.g., when...). >0.7, selected as the i-th parent) to generate hybridization combinations, with fresh grass yield, crude protein content and salt tolerance as the core evaluation criteria, and constructing an objective function: ; Step S2.3: Dynamic simulation of the hybridization process By incorporating parental gene expression data, simulating gene recombination behavior during hybridization of different strains, and predicting the phenotypic probability distribution of offspring, a dynamic hybridization simulation model is formed. ; Let the expected value of the sub-representative of the k-th combination be . Let be the expression level of the gene associated with the j-th target trait. is the weight factor of the j-th gene in the k-th combination, obtained by fitting a linear regression model of parental gene expression data. The model training uses the least squares method. g is the number of target genes participating in the simulation. The summation sign of subscript j from 1 to g represents the summation of the weighted expression values ​​of all genes related to the target trait. E is the mathematical expectation operation, which represents the weighted average of all possible offspring gene combinations under the condition of uncertainty in gene expression and recombination.

[0040] Step S2.4: Evolutionary Process and Dynamic Parameter Adjustment During the algorithm iteration process, a dual-elite mechanism and an adaptive parameter adjustment strategy are adopted; Elite replication: Preserving fitness before each generation Individuals ( (Initial value: 0.15) Mutation operation: generating Random chromosomes to increase diversity ( (Initial value: 0.05) Bias crossover: Selecting elite genes with a probability p to generate the next generation of individuals. ; in, This represents the value of the i-th gene in the offspring, specifically the coding value at the i-th position in the new individual generated after mating. It is a component of the next generation's chromosome. The gene originating from elite parent 1 indicates that the gene locus was selected from individuals with high fitness (i.e., good performance) in the current population. Let be the gene from the non-elite parent 2, and let represent the gene locus selected from the non-optimal individuals in the current population. p is the probability of retaining elite genes, indicating the probability that the i-th gene will be directly inherited from the elite individual when generating a new generation of chromosomes. Since p > 0.5, this strategy has a clear bias towards elites, strengthening the transmission of superior genes. 1-p is the probability of non-elite genes being selected, used to preserve population diversity and avoid getting trapped in local optima. Optionally, the bias probability p increases linearly from 0.6 to 0.8.

[0041] As the number of iterations progresses, the system dynamically adjusts parameters such as the elite ratio and mutation probability to adapt to the search progress, as shown in the following formula: ; Let be the proportion of elite individuals at the t-th algorithm iteration. The maximum number of iterations, The minimum and maximum elite ratio thresholds are set. The dynamic adjustment range is: .

[0042] Step S2.5: Jitter and Reset Mechanism To avoid the algorithm getting trapped in local optima, by introducing... Proportional mutant individuals (completely randomly generated new solutions) enhance population diversity; elite individuals are slightly perturbed using a jitter operator; elite individuals are probabilistically... Apply a disturbance; When there is no improvement in fitness for several consecutive generations, a reset operation is performed to remove non-elite individuals and regenerate new chromosome individuals in order to avoid premature convergence. Step S3: Through transcriptome-metabolome-molecular marker joint analysis, construct a gene-metabolome-phenotype association model to achieve early screening of hybrid offspring. Step S3 constructs a gene-metabolome-phenotype association model by analyzing the nonlinear relationship between target gene expression level, key metabolite content and phenotype (fresh grass yield, crude protein content), and achieves early screening of hybrid offspring based on this model, with fresh grass yield and crude protein content as the dual superparents as screening targets.

[0043] Step S3 involves the specific process of early and precise screening of hybrid offspring, including: Step S3.1: The step S3.1, which constructs a multi-omics association model, involves building an association model through joint analysis of transcriptomics, metabolomics, and molecular markers. This includes screening differentially expressed genes related to yield and quality through transcriptomics sequencing and detecting the content of key metabolites through metabolomics. A hybrid dataset was constructed by combining SSR molecular marker data. Correlation analysis was used to fit the gene-metabolite-phenotype associations, resulting in a gene-metabolite-phenotype association model. Threshold screening and marker validation were then used to establish an early screening mechanism. This model efficiently simulates the phenotypic performance of hybrid offspring. Specifically, the input data consisted of differentially expressed genes from the transcriptome. Key metabolome data and molecular marker data The training set is composed of various elements. .

[0044] By pooling data, the strong associations between differentially expressed genes and metabolites are leveraged to narrow the screening pool, avoiding the inefficiency of random screening. Simultaneously, molecular marker data is used to correct model errors, enhancing adaptability to genetic background differences, improving screening generalization, and preventing model overfitting to single datasets. Transcriptome differentially expressed gene data are also utilized. , This includes expression data for the RUBISCO gene (fold change = 3.2) related to fresh grass yield and the GS1 gene (fold change = 2.8) related to crude protein synthesis, as well as key metabolome data. Includes: content data of free amino acids (12-25 mg / g) and soluble sugars (8-18 mg / g), and molecular marker data. This includes the detection results for SSR markers Xmsat106 and Xmsat448.

[0045] Correlation analysis (Pearson correlation coefficient) was used to fit the gene-metabolite-phenotype association. The association model was defined with gene expression level G and metabolite content M as inputs and phenotypic prediction value P (fresh grass yield / crude protein content) as output. The selection criteria were gene-phenotype correlation coefficient |R| ≥ 0.7 and metabolite-phenotype correlation coefficient |R| ≥ 0.6. The training objective was to minimize the mean squared error between the predicted phenotype and the true phenotype. The loss function was adjusted by adding weight decay to avoid overfitting. ; in, This is the model loss function, used to measure the gap between the current learned model and the true correlation. This is a hybrid dataset, a joint dataset composed of transcriptomic, metabolomic, and molecular marker data. These are the true phenotypic values. The phenotypic values ​​predicted by the model. For model parameters, Regularization coefficient It is the square of the L2 norm, i.e., the regularization term, used to constrain the parameter size.

[0046] This loss function can prevent the model from relying too much on a single data distribution and enhance its adaptability to samples with different genetic backgrounds.

[0047] Step S3.2: In-model screening and optimization. Step S3.2 involves screening core genes from the differentially expressed gene set as initial targets for early screening. Within the association model, a threshold method is used to screen hybrid offspring that meet the target trait. The screening results are then validated and optimized using molecular markers. The resulting compliant lines are further fed back into the hybridization process to generate new data. This forms a data closed loop. Specifically, core genes (RUBISCO, GS1) are screened from the differential gene set as initial targets for early screening, simulating the starting point for efficient screening. In the learning model, the screening strategy is optimized for the current screening threshold to maximize the proportion of compliant strains. ; in, Define the screening rules for a given gene-metabolite data as a screening strategy; This is the new selection strategy obtained by optimization at iteration t+1. To filter trajectories From the form {( ),...,( The gene-metabolite-phenotype pair sequence structure of )} is constructed through a strategy In the association model The number of qualified strains, i.e., the strain coefficient that meets the criteria of "fresh grass yield exceeding parentage rate ≥10% and crude protein content ≥20%", is generated by simulated sampling. To optimize the objective, consider all possible screening strategies. The strategy that maximizes the number of compliant strains is chosen, i.e., the optimal screening strategy is found.

[0048] Thresholding or cluster analysis methods can be used to efficiently iterate screening strategies in the model, reducing reliance on real phenotypic measurements. Low-cost iterative strategies can be implemented in the virtual model to reduce the number of real tests. The screening threshold can be used to dynamically guide the strategy to approach the optimal standard.

[0049] Step S3.3: Filtering Strategy Update. Optimize the filtering threshold using feedback data to gradually improve filtering accuracy. ; in, This is the current filtering threshold, used in the t-th iteration; Let be the filtering threshold after the (t+1)th iteration. The adjustment rate is used to control the update step size of the threshold.

[0050] Optionally, based on preliminary experimental verification, when the screening strategy... If the error rate is below 5% for five consecutive iterations, and further optimization contributes less than 1% to the improvement of the pass rate, then the optimization is terminated.

[0051] By continuously refining the screening strategy, the model screening results are made to increasingly approximate the true phenotype, thus realizing a closed-loop learning process of screening strategy updates, model corrections, and screening optimization.

[0052] The optimized screening strategy was applied to the screening of real hybrid offspring, and new phenotypic data were collected. The new data underwent validity screening, retaining data with high correlation and high diversity for further training of the model and screening strategy. The final experimental results determined the optimal screening strategy: RUBISCO gene expression fold change ≥3.0, GS1 gene expression fold change ≥2.5, free amino acid content ≥20 mg / g, SSR marker double positivity, and prediction achievement rate ≥85%.

[0053] Step S3 constructs a multi-omics association model, nesting and iterating gene screening, metabolite detection, and molecular marker verification to achieve low-cost screening in a virtual environment (reducing the need for real phenotypic determination), self-correcting screening strategies (avoiding the subjectivity of experience-based screening), and deep integration of multi-omics data with the association model to form a self-iterable screening closed loop, thus enabling early and accurate screening of alfalfa hybrid offspring. Step S4: Construct an environmental dynamic model based on a long short-term memory neural network, and search for the optimal cultivation strategy that maximizes comprehensive traits in a virtual environment. This step involves analyzing the iron ion concentration in the culture medium (range: 0–2.5 g / kg) and the fresh fruiting body yield (range: 19.2–20.5 t / hm²). 2 The nonlinear relationship between the crude protein content (range: 20.1–22.3%) and the environmental dynamic model was constructed. Based on the long short-term memory neural network, the optimal cultivation formula was automatically searched to maximize the fresh grass yield and achieve a crude protein content ≥20%.

[0054] Step S4, which implements intelligent prediction of culture conditions, includes the following specific processes: Step S4.1: Constructing a Long Short-Term Memory Neural Network Environment Model. This step constructs a dynamic environment model using a long short-term memory neural network, including: constructing a training dataset using historical training data and real-time monitoring data; using a long short-term memory neural network to fit the mapping relationship between environmental factors and traits to obtain a dynamic environment model; and using strategy optimization and prediction function iteration to form a closed-loop learning mechanism to efficiently simulate the adjustment behavior of training conditions in a virtual model environment. Specifically: the input data consists of historical high-quality training data. (High-quality priors) and real-time monitoring data (Real-time correction), mixed to form the training set: ; By mixing data, high-quality environment-trait pairs are provided using historical high-quality data, narrowing the policy search space and avoiding the inefficiency of random exploration. Simultaneously, real-time monitoring data is used to correct model errors, enhancing adaptability to environmental fluctuations, improving the generalization of dynamic environmental predictions, and preventing model overfitting to historical data. Historical high-quality training data is also utilized. This includes: artificially optimized culture medium formula (such as ferrous sulfate 1.5–2.5 g / kg), temperature and humidity control curve (20–25℃ during the seedling stage), and harvest time (30 days after flowering).

[0055] A long short-term memory neural network is used to simulate environmental dynamics, defining the state space S = {culture medium formula C, temperature and humidity T, iron ion concentration Fe, mycelial growth rate G}, and the action space A = {adjusting Fe concentration}. Change temperature Modify the formula ratio The prediction function f(s,a) is used to predict the phenotypic values ​​(fresh grass yield Y, crude protein content CP) after performing action a in state s, where Y is the measured fresh grass yield and CP is the measured crude protein content. The model weights can be dynamically adjusted with each iteration stage, initially focusing on exploration (uniform weight distribution) and later focusing on convergence (yield weight increased to 0.4), predicting the next state distribution and phenotypic values ​​after performing action a in state s. The training objective is to minimize the mean squared error between the predicted phenotypic and the true phenotypic, and the loss function is... Overfitting can be avoided by adding weight decay and gradient clipping: ; in, This is the model loss function, used to measure the difference between the current learned model and the dynamics of the real environment. For a mixed dataset, consisting of historical data and real-time data The dataset consists of two parts: s represents the current state, and a represents the action. , This represents the actual fresh grass yield and crude protein content. , For the prediction of fresh grass yield and crude protein content, Here, λ is the model parameter, and λ is the regularization coefficient. It is the square of the L2 norm, i.e., the regularization term, used to constrain the parameter size.

[0056] This loss function can prevent the model from relying too heavily on historical data distribution and enhance its adaptability to complex environments with small samples.

[0057] Step S4.2: Intra-model Strategy Search. Step S4.2 extracts environmental parameter combinations from historical high-quality cultivation data as the initial state for strategy optimization. Within the dynamic environmental model, the gradient descent method is used to optimize the cultivation strategy. A new strategy is generated to maximize the overall score and update the prediction function through gradient ascent. This data is further fed back into the real culture environment, generating new data. This forms a closed data loop.

[0058] Specifically: Randomly sample states from historical data This serves as the initial state for strategy optimization within the model, simulating the starting point for high-quality cultivation behavior.

[0059] In the learning model Ŵ, for the current prediction function Optimization strategy Maximize the overall score: ; in, To cultivate a policy, a probability distribution for choosing an action in a given state is defined. This is the new strategy obtained in the (t+1)th iteration. , For trajectory From the form {( ), ......, ( The state-action pair sequence is constructed using a policy. In environmental dynamics model Score (Simulated sampling generation) For the state at the h-th step of the trajectory and actions The overall score.

[0060] Gradient descent or model predictive control methods can be employed to implement efficient iterative strategies within the model, reducing frequent reliance on the real environment. Low-cost iterative strategies in a virtual environment reduce the number of real experiments and utilize prediction functions. Dynamic guidance strategies approximate optimal cultivation conditions.

[0061] Step S4.3: Prediction Function Update. The prediction function is optimized through gradient ascent to perform regretless learning, gradually improving prediction accuracy. ; in, This is the current prediction function, used in the t-th iteration. Let be the prediction function after the (t+1)th iteration update. The learning rate controls the update step size of the prediction function. To find the gradient of the prediction function f, In the prediction function Below, the loss between actual output and projected output, For the prediction function The difference between actual and projected output is then considered.

[0062] Optionally, based on preliminary experiments, if the loss of the prediction function f increases by less than 1% over 5 consecutive iterations, and further optimization contributes less than 0.5% to the phenotypic improvement, then the optimization is terminated.

[0063] By continuously refining the prediction function, the training strategy is made to approach the optimal conditions, thus realizing a closed-loop reinforcement learning process of prediction function update, strategy optimization, and model correction.

[0064] The optimized strategy was applied to a real-world training system to collect new trajectory data. The new data underwent validity screening, retaining high-scoring and highly diverse trajectories for further training of the model and prediction function. The final output results obtained through the experiment are as follows: Recommended optimal cultivation formula: 50% corn stalks + 30% soybean meal + 15% wheat bran + 5% inorganic salt mixture + 2.0 g / kg ferrous sulfate (60% moisture content), with a predicted fresh forage yield of 20.5 t / hm². 2 The crude protein content is 22.1%. Recommended environmental parameters: seedling temperature 20-25℃, flowering temperature 25-28℃, and relative humidity 60-70%.

[0065] Step S4 constructs a long short-term memory neural network environment model, nesting and iterating the three processes of environment modeling, strategy optimization, and prediction update. This enables low-cost trial and error in the virtual environment (reducing real breeding experiments), self-correction of the prediction function (avoiding the subjectivity of manually designed cultivation conditions), and deep integration of the long short-term memory neural network with the dynamic environmental model to form a self-iterative decision-making closed loop, thus realizing intelligent prediction and autonomous optimization of alfalfa cultivation conditions. Step S5: Construct a multi-index comprehensive scoring system based on the analytic hierarchy process (AHP), identify abnormal samples and trigger model backtracking adjustments, and output high-performance strains to enter the next round of breeding.

[0066] Step S5 comprehensively evaluates the performance of candidate strains in terms of fresh grass yield, crude protein content, and salt tolerance (growth period), screens high-performance strains, and backtracks to adjust the model strategy when standards are not met, thereby improving breeding efficiency. Step S5, the process of strain performance evaluation and iteration, specifically includes: Step S5.1: Construction of a multi-indicator comprehensive evaluation system. The analytic hierarchy process (AHP) is used to assign weights to each indicator: fresh grass yield 35%, crude protein content 30%, salt tolerance 20%, and growth period 15%. The comprehensive score for each strain is calculated as follows: The comprehensive scoring system formula is: ; in This represents the actual yield of fresh grass. This represents the minimum yield of fresh grass. This represents the maximum yield of fresh grass. This represents the maximum crude protein content. This represents the minimum crude protein content. This is the measured value of crude protein content. This is the measured value of salt tolerance. This is the optimal value for the reproductive period. These are measured values ​​during the reproductive period.

[0067] Step S5.2: Automatic Identification and Removal of Abnormal Samples The following types of samples will be identified as abnormal and automatically removed: Fresh grass yield was lower than the average level of the parent plants; Crude protein content is less than 20%; Abnormalities during the reproductive period; The logic for removing samples with missing indicators or obvious data errors is as follows: if the fresh grass yield is less than the average value of the parent strain or the crude protein content is less than 20%, the sample of that strain will be removed.

[0068] Step S5.3: Iterative Optimization and Model Backtracking Adjustment If most strains fail to meet the standards in a certain round of evaluation (e.g., the overall score is below 60 points or the crude protein content is below the standard), the system will automatically backtrack to the previous stage (e.g., parental selection or hybridization design) and perform the following optimizations: Increase the weight of high-yield, high-crude-protein markers to enhance the selection bias for target traits; Re-screen the parents and remove materials with excessively low basic traits; Update the parameters of the genetic algorithm or association model to optimize the prediction quality of hybridization combinations.

[0069] Parameter adjustment logic can be expressed by adding a penalty term (which only takes effect when the core trait is not met). That is, if the strain's compliance rate is lower than the threshold, the weight of the core trait is increased or a penalty term is added. The total loss is calculated as follows: ; The penalty coefficient is dynamically adjusted based on the degree to which the core trait is not met.

[0070] According to step S5.3, when the overall score of the strain is lower than the preset threshold or the core traits are not up to standard, the system automatically backtracks to the parent selection or hybridization design stage and updates the parameters of the genetic algorithm or association model to optimize the prediction quality of hybridization combinations.

[0071] Step S5.4: Output and Next Round Evaluation. The top 10% of strains in the overall score are retained as excellent candidates and enter the next round of experiment; other samples are used for model training or elimination. The system records and outputs the following information: the score distribution map of each round, the list of eliminated samples and reasons for elimination, and the adjustment record of model parameters and weights, which is convenient for tracking the optimization path.

[0072] Furthermore, based on the above-mentioned screening method for high-quality and high-yield alfalfa hybrid combinations integrating multiple technologies and the optimal cultivation strategy determined by the system, the present invention conducted breeding experiments, and the experimental results are as follows: The first round of creation of high-quality, high-yield hybrid strains: using high-yield alfalfa cultivation strain B218 (fresh grass yield 17.8 t / hm). 2 Crude protein content 18.5% and high crude protein alfalfa wild strain C356 (fresh grass yield 15.2 t / hm) 2Using a parent plant (with a crude protein content of 20.1%), homozygous strains were obtained by isolating individual plants after seed harvest. After verification of homozygous strain authenticity (SSR marker detection), 18 homozygous strains were prepared from the two parents. These were then hybridized, resulting in 81 hybrid combinations. After SSR marker verification, 36 true hybrid strains were screened out for field trials. The fresh grass yield and crude protein content were measured (see Table 1). The D48 strain, which exhibited both high fresh grass yield and high crude protein content, had a fresh grass yield of 17.9 t / hm². 2 The crude protein content is 20.8%; Table 1 Fresh grass yield and crude protein content of the first round of hybrid strains

[0073] 1. Second round of creation of high-quality and high-yield hybrid strains: Using the selected high-quality and high-yield hybrid strain D48 as the parent, seeds were harvested to obtain homozygous strains. After authenticity identification of the homozygous strains, a total of 6 homozygous strains were prepared. Backcrossing experiments were conducted with the parent C356-4 to obtain 8 hybrid strains. After verification by SSR markers, 5 genuine hybrid strains were screened out for field trials to detect fresh grass yield, crude protein content, and agronomic traits (Tables 2 and 3). Finally, hybrid strain JAUA086 was selected as having the highest fresh grass yield and crude protein content, with a fresh grass yield of 19.2 ± 0.6 t / hm. 2 The crude protein content was 22.3±0.4%, with a super-high content of 10.9%. Table 2 Fresh grass yield and crude protein content of the second-round hybrid strains

[0074] Table 3 Agronomic traits of hybrid strains

[0075] 2. Screening of culture medium formula: 1) Experimental treatment: The iron ion concentration gradient of the culture medium was set at 1.0 g / kg, 1.5 g / kg, 2.0 g / kg and 2.5 g / kg. Different concentrations of ferrous sulfate solution were added to the traditional culture medium formula (50% corn stalks, 30% soybean meal, 15% wheat bran and 5% inorganic salt mixture). The specific iron ion concentration gradient of the culture medium is shown in Table 4. The corresponding concentration of ferrous sulfate was weighed and dissolved in the water for mixing. After it was completely dissolved, it was mixed into the culture medium, mixed evenly and then put into a pot.

[0076] Table 4. Details of Iron Ion Addition at Different Concentrations

[0077] 2) Cultivation and Management: Use plastic pots (30cm in diameter, 25cm in height) for cultivation. Fill each pot with 5kg of substrate, autoclave for 2 hours, and sow after the substrate has cooled to room temperature. The sowing rate is 15g per pot. During cultivation, maintain a seedling temperature of 20-25℃ and a flowering temperature of 25-28℃, with a relative humidity of 60-70%. Water regularly to maintain a moisture content of 60%.

[0078] 3) Characteristic determination: Harvest 30 days after flowering and determine the yield of fresh grass; determine the crude protein content using the Kjeldahl method.

[0079] 4) Screening of culture medium formula for strain JAUA086: Based on the fresh grass yield and crude protein content of different treatment groups and control groups, the optimal iron ion concentration in the culture medium was selected. The results showed that for strain JAUA086, the fresh grass yield first increased and then decreased with increasing iron ion concentration in the culture medium, reaching its highest level of 20.5 ± 0.5 t / hm² at a concentration of 2.0 g / kg, while the crude protein content remained at 22.1 ± 0.3% (Table 5). Therefore, the optimal culture medium formulation for strain JAUA086 is: 50% corn stalks, 30% soybean meal, 15% wheat bran, 5% inorganic salt mixture, 2.0 g / kg ferrous sulfate, and 60% moisture content.

[0080] Table 5. Fresh grass yield and crude protein content of strain JAUA086

[0081] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for screening high quality and high yield alfalfa hybrid combinations by fusing multiple technologies, characterized in that, The method comprises the following steps: Step S1: Collecting alfalfa strain basic data, hybridization process data, culture environment parameters and external reference data, and adopting standardized score normalization and one effective code to perform standardization and structured processing; Step S2: Using the improved non-dominated sorting genetic algorithm-II to perform multi-objective coding optimization on the hybrid combination, taking fresh grass yield, crude protein content and salt tolerance as core targets to construct a fitness function, and combining a double-elite reservation mechanism and an adaptive crossover and mutation strategy to generate an optimal hybridization scheme; Step S3: Constructing a gene, metabolite and phenotype correlation model through transcriptome differential analysis, metabolome key substance detection and simple sequence repeat molecular marker verification, and realizing early screening of hybrid offspring; Step S4: Constructing an environmental dynamic model based on a long short-term memory neural network, simulating strain growth performance under soil water content, salinity and temperature gradient, and searching for an optimal environmental regulation strategy maximizing comprehensive traits; Step S5: Constructing a multi-index comprehensive scoring system based on an analytic hierarchy process, assigning weights to fresh grass yield, crude protein content, salt tolerance and growth period, identifying abnormal samples and backtracking to adjust parameters of the previous model, and outputting a high-performance hybrid combination.

2. The method for screening high quality and high yield alfalfa cross combinations by fusing multiple technologies according to claim 1, characterized in that, In the step S1: Strain basic data collection: collecting genetic background data, agronomic trait data and stress resistance data of alfalfa parents; The hybridization process data includes single hybrid seed setting rate, generation plant coefficient, generation segregation ratio; Culture environment parameters include drought stress gradient, salinity gradient and temperature; External reference data include expression patterns of Medicago trunculata and stress resistance genes and multi-trait breeding cases of legume plants.

3. The method of claim 1, wherein the method is a method of screening for high quality high yielding alfalfa cross combinations using multi-technologies, characterized by, In the step S2: The hybrid combination is coded as a 12-dimensional continuous real number represented chromosome vector X; The chromosome vector X is decoded into a specific combination strategy and fitness score by an encoder, and a target function is constructed taking fresh grass yield over-parent rate, crude protein content and salt tolerance as evaluation cores; Parent gene expression data is introduced to simulate gene recombination behavior when different strains are hybridized, predict trait probability distribution of offspring, and form a dynamic hybridization simulation model; Employing a dual-elite evolutionary mechanism, each generation retains the previous generation's fitness. A select few individuals, and with probability >0.5 Perform biased crossover operation, and offspring genes are distributed with probability. Inheriting the elite parent's gene 1 with a probability of 1- Inherited two genes from non-elite parents; With the advancement of iteration number, the elite ratio and mutation probability are dynamically adjusted to adapt to the search progress.

4. The method for screening high quality and high yield alfalfa cross combination of fusion multi-technology according to claim 3, characterized in that, The target function in the step S2 is: ； wherein, represents the fitness score corresponding to chromosome X, represents the fresh grass yield over-parent rate, i.e. the ratio of the predicted offspring fresh grass yield being higher than the average of the parents, represents the crude protein content normalized value, represents the salt tolerance index, , , represents the objective function weight parameter, which is set according to the task requirement; The improved non-dominated sorting genetic algorithm-II also comprises: Shaking mechanism: perturbing elite individuals with probability Applying perturbation; Reset mechanism: when there is no fitness improvement for a continuous preset number of generations, the non-elite individuals are removed and random chromosomes are generated again.

5. The method for screening high quality and high yield alfalfa cross combinations by fusing multiple technologies according to claim 1, characterized in that, The step S3 constructs a correlation model through transcriptome, metabolome and molecular marker joint analysis, comprising: Differential genes related to yield and quality are screened through transcriptome sequencing, key metabolite content is detected through metabolome detection, and a mixed data set is constructed combining SSR molecular marker data; The correlation between genes, metabolites and phenotypes is fitted through correlation analysis to obtain a gene, metabolite and phenotype correlation model, and an early screening mechanism is formed through threshold screening and marker verification to efficiently simulate trait performance of hybrid offspring in the model.

6. The method for screening high quality and high yield alfalfa cross combination of fusion multi-technology according to claim 1, characterized in that, The step S3 also comprises: Core genes are screened from the differential gene set as initial target points for early screening; Within the association model, a threshold method is used to screen hybrid offspring that meet the target trait, and the screening results are optimized through molecular marker validation. The generated compliant lines are further fed back into the hybridization process to produce new data. This forms a closed data loop.

7. The method for screening high quality and high yield alfalfa cross combination of fusion multi-technology according to claim 1, characterized in that, In the step S4: Long short-term memory neural network: the input of the environmental dynamic model is soil water content, sodium chloride concentration and temperature, and the output is predicted values of fresh grass yield and crude protein content; Environmental dynamic model: trained by multi-region field data to realize accurate prediction of strain growth performance under different environments; The optimal environmental regulation strategy is the regulation and change of soil water content, salinity and temperature in different periods of seedling stage.

8. The method of claim 1, wherein the method is a method of screening for high quality high yielding alfalfa cross combinations using multi-technologies, characterized by, In the step S5, The comprehensive scoring system formula is: ； wherein is the actual value of fresh grass yield, is the minimum value of fresh grass yield, is the maximum value of fresh grass yield, is the maximum value of crude protein content, is the minimum value of crude protein content, is the actual value of crude protein content, is the actual value of salt tolerance, is the optimum value of growth period, is the actual value of growth period; Identify abnormal samples: trigger backtracking adjustment mechanism for abnormal samples to adjust the parameters of the previous screening link to optimize the subsequent screening results.

9. The method of claim 1, wherein the method is a method of screening for high quality high yielding alfalfa cross combinations using multi-technologies, characterized by, In the step S5, Each score is standardized. When the comprehensive score of the strain is lower than the preset threshold or the core traits do not meet the standards, the system automatically backtracks to the parent selection or hybridization design link and updates the parameters of the genetic algorithm or the correlation model to optimize the prediction quality of the hybrid combination.

10. A multi-technology integrated high quality and high yield alfalfa hybrid combination screening system, characterized in that, It includes: Data processing module: used for collection, standardization and structured processing of multi-source data; Hybrid optimization module: running an improved multi-objective genetic algorithm to generate an optimal hybrid scheme; Multi-omics screening module: integrating multi-omics and molecular marker data to realize early screening; Environmental simulation module: based on recurrent neural network to build dynamic model and output environmental regulation strategy; Performance evaluation module: output high-performance hybrid combination through hierarchical evaluation and backtracking adjustment.

Images