Intelligent navigation breeding qtn data analysis method based on large model

Abstract

The application relates to the field of rice breeding, and specifically discloses an intelligent navigation breeding QTN data analysis method based on a large model, which comprises the following steps: constructing a structured QTN-phenotype-genetic effect value database, performing verification data retrieval on each QTN site based on a literature database, extracting planting location and environment information contained in the verification data and phenotype information disclosed under the planting environment, determining the standardization effect value of each QTN site, constructing an LSTM network based on a fusion vector, obtaining a trained LSTM network based on the fusion vector, extracting QTN sites of a to-be-predicted variety and target planting environment information, and inputting the extracted QTN sites of the to-be-predicted variety and the target planting environment information into the trained LSTM network based on the fusion vector for prediction.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of rice breeding technology and artificial intelligence, and in particular to an intelligent data analysis method that combines neural networks and rice genomics. Background Technology

[0002] Currently, high-throughput sequencing technology is widely used in rice molecular breeding, enabling rapid and low-cost acquisition of whole-genome quantitative trait nucleotide (QTN) data for breeding materials, covering millions to tens of millions of genetic variation sites. These genetic markers provide an unprecedented data foundation for analyzing the genetic structure of important agronomic traits such as yield, quality, and stress resistance in rice, locating key functional sites, and implementing marker-assisted selection (MAS) or genome selection (GS). Mainstream breeding analysis platforms mostly employ statistical prediction models based on population genetic covariance structures, such as Genomic Best Linear Unbiased Prediction (GBLUP), Bayesian regression series (BayesA / B / C, etc.), or Ridge Regression (RR-BLUP), to perform multi-dimensional phenotypic predictions of yield components such as number of grains per panicle, thousand-grain weight, and effective tiller number; rice quality indicators such as amylose content, chalkiness, and gelatinization temperature; and biotic and abiotic stress resistance traits such as rice blast resistance, drought resistance, salt tolerance, and lodging resistance. To help users understand the analysis results, some platforms also provide basic visualization functions, which improves the intuitiveness of data interpretation to some extent.

[0003] However, the aforementioned technical approaches are essentially still based on a linear model framework dominated by "additive genetic effects." Their underlying mathematical assumptions are ill-suited to the widespread nonlinear genetic mechanisms present in the complex traits of rice, including epistatic interactions between genes, dominant effects, and dynamic interactions between genotypes and environmental factors (such as temperature, water, and fertilization levels) and even cultivation techniques. This results in poor predictive stability of the models across environmental and population scenarios, with systematic deviations often appearing between field measurements and model predictions. Furthermore, as cultivation techniques, pest and disease control technologies, and fertilizer fertility improve, the planting effects of the same variety evolve over time. Existing models are entirely based on calculations and simulations of the characteristics of individual gene loci, making it difficult for breeders to formulate precise gene editing or parental selection strategies. Summary of the Invention

[0004] During the research and development process, the inventors of this application noticed that although the amount of QTN data was sufficient and could be used to predict phenotypes, such predictions often deviated significantly from field planting results. Furthermore, these differences failed to reflect the evolution of cultivation techniques and scientific management methods; that is, even the same rice variety yielded significantly differently twenty years ago compared to today. Moreover, the lack of data from field planting experiments resulted in a mismatch between QTN and field planting data, making it difficult to obtain a predictive model that incorporates both field planting environment data and QTN data.

[0005] After conducting extensive simulations and analyses, the inventors of this application discovered that, in order to accurately predict complex quantitative traits of rice (such as yield, quality, and stress resistance), it is necessary to consider the influence of planting time and planting environment on the prediction, and to effectively integrate and utilize massive amounts of genomic data (QTN loci), planting environment data, genetic effect information, and domain expertise.

[0006] Therefore, this invention proposes to first construct a structured processing of QTN site data and introduce the concept of a comprehensive effect value. A comprehensive evaluation of the QTN data is first formed using a large amount of QTN data. Then, the obtained comprehensive effect value is fed into an LSTM network based on fusion vectors to make actual predictions of phenotypes based on environmental data.

[0007] Specifically, this invention provides a QTN data analysis method for intelligent navigation breeding based on a large model, the method comprising:

[0008] Step (1) Construct a structured QTN-phenotype-genetic effect value database, which includes the ID, phenotypic information, and genetic effect values ​​of each QTN locus in previous experiments. Step (2) Based on experimental data and literature databases, perform validation data retrieval for each QTN locus, extract the planting location and environmental information and the phenotypic information under the disclosed planting environment contained in the validation data, add time tags to all data, and sort them in chronological order to form a time data sequence. Step (3) For each QTN locus in the database, average the genetic effect values ​​of each experiment containing that locus to obtain the value of each QTN locus. The standardized effect value of the point; Step (4) Construct an LSTM network based on fusion vectors. This LSTM network based on fusion vectors receives gene information and environmental information respectively. The gene information includes the genotype information, phenotypic information and standardized effect value of the QTN loci contained in the variety or gene to be predicted. The environmental information includes location and planting condition information. The planting location and environmental information contained in the time data sequence and the phenotypic information in the disclosed planting environment are used as data to train the network and obtain the trained LSTM network based on fusion vectors. Considering that planting is basically carried out on an annual basis, the time data sequence is sorted by year and the same weight value is assigned to the same year. Step (5) Extract the QTN loci of the variety to be predicted and the target planting environment information. The extracted QTN loci contained in the variety to be predicted and the target planting environment information are converted into fusion vectors and input into the trained LSTM network based on fusion vectors for prediction.

[0009] Preferably, the initial genetic effect value is calculated according to the following formula. , where n represents the number of statistical trials in a single batch, and P represents the genetic effect value of a single batch (or prediction). This represents the measured value of a certain trait in non-wild-type (mainly mutant) plants. This represents the measured value of wild-type plants for the same trait.

[0010] Preferably, for the gene information input branch, each parameter included is used as a dimension of a vector to construct a first multidimensional vector X. t1 ={X 1， X 2， X 3，…. X S1 For the planting environment input branch, each planting environment-related parameter is used as one dimension of a vector to construct a second multi-dimensional vector, X. t2 ={X S1+1， X S1+1， X S1+1，…. X S1+S2The second multidimensional vector is multiplied by a weighting coefficient, which represents the improvement in cultivation techniques, and then concatenated with the first multidimensional vector to obtain the concatenated input vector X. t ={ X 1， X 2， X 3，…. X S1, λ1(t)X S1+1， λ2(t)X S1+2， λ(t)X S1+3，…. λ S2 (t)X S1+S2}, where S1 represents the dimension of the first multidimensional vector, S2 represents the dimension of the second multidimensional vector, λ1(t), λ2(t), ..., λ S2 (t) represents the weight of each dimension in the second multidimensional vector.

[0011] Preferably, each unit of the LSTM network includes an input gate i(t), an output gate O(t), and a forget gate f(t).

[0012] Preferably, the database contains at least 320 QTN loci.

[0013] Preferably, the associated traits include one or more of the following: disease resistance, drought tolerance, salt tolerance, nitrogen uptake efficiency, lodging resistance, amylose content, gelatinization temperature, thousand-grain weight, seed setting rate, seed coat color, thousand-grain weight, heading date, plant height, and rice quality characteristics.

[0014] Preferably, the genetic information includes parameters such as gene ID, gene name, chromosome, associated traits, wild-type / mutant phenotype, and standard genetic effect value.

[0015] Preferably, the planting environment parameters include one or more of the following: temperature, light duration, rainfall during the growing season, rainfall during the maturity period, and fertilization status.

[0016] Preferably, step (3) further includes constructing a convolutional neural network model (CNN model), using the QTN locus ID, phenotypic information and genetic effect values ​​of previous experiments in the database as training data, using the QTN loci and corresponding phenotypic information contained in the database as input, and using the standardized effect values ​​of the corresponding phenotypic information as output information to train the first convolutional neural network model to obtain the standardized effect value of each QTN locus.

[0017] On the other hand, the present invention provides an application of the above method, the application including one of the following:

[0018] (1) To more accurately predict various traits of the target variety and to accurately locate the target points that need genetic improvement for the recipient parent;

[0019] (2) To enable intelligent screening of optimal donor parents for molecular breeding;

[0020] (3) Provide intelligent navigation support for high-throughput screening and targeted selection of future generations of materials;

[0021] Prior to the method proposed in this invention, existing rice molecular breeding phenotypic prediction mainly relied on traditional statistical models. These models assumed that it was difficult to effectively use the improvement of cultivation techniques and scientific management methods over time as auxiliary criteria for quantitative trait genes. They also failed to capture non-additive genetic mechanisms such as epistasis, dominance effects, and gene-environment interactions, resulting in a lack of biological interpretability in the predictions.

[0022] This invention effectively integrates previously difficult-to-match genetic information and planting information by using genetic effect values ​​as a link. It can make full use of the research results on gene locus traits in gene locus-related studies and the influence of planting time and geographical environment on planting results, especially quantitative traits, in actual planting studies. It can more effectively combine the research results on gene locus traits with the research results on quantitative traits in planting, so as to achieve prediction results that are closer to the actual situation. Furthermore, it can provide more reasonable planting suggestions based on the target planting area, creating a closed loop of intelligent breeding across the entire chain.

[0023] This invention proposes a large-scale model-based navigation breeding QTN data analysis method and system, and constructs an end-to-end intelligent analysis system covering "gene sequencing input → multidimensional trait prediction".

[0024] This invention adopts a prediction paradigm of "effect value simulation + actual planting environment enhancement", and makes accurate biologically interpretable estimations of key quantitative traits based on the genetic effect parameters and the genotype-phenotype mapping relationship. Attached Figure Description

[0025] Figure 1 This is a flowchart of the QTN data analysis method of the present invention;

[0026] Figure 2 This is a schematic diagram of the neural network structure used in the method of the present invention;

[0027] Figure 3 This is an example of classifying and labeling different QTN sites. Detailed Implementation

[0028] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0029] Example 1: A QTN data analysis method based on a structured knowledge base is used to intelligently predict key agronomic traits of rice through integrated analysis of QTN data.

[0030] This embodiment provides an intelligent prediction method for key agronomic traits of rice based on a structured genetic knowledge base. Specifically, it includes the construction of a rice quantitative trait nucleotide (QTN) locus database, sample genotype identification and phenotypic mapping (preferably also including redundant phenotypic deduplication), and trait prediction and risk assessment based on QTN effect values. This embodiment uses rice sample number "ZKW24001103" as an example for detailed explanation.

[0031] Step 1.1: Construct a structured QTN-phenotype-effect size knowledge base

[0032] The system collects 320 core functional QTN loci verified by publicly published literature or authoritative databases to construct a structured database. Each locus includes a unique ID, gene name, chromosome, associated trait, and wild-type / mutant phenotypic description. For continuous traits (such as heading date and plant height), the system assigns a genetic effect value to each reported data point based on the degree of influence mentioned in relevant reports. If the degree of influence is not recorded in the report, a default genetic effect value is set or the data is discarded. The phenotypic effect of a specific QTN locus depends on its genotype (mutant or wild-type) and the external conditions. Therefore, the agronomic trait phenotypes corresponding to favorable and unfavorable alleles (alleles to be improved) are labeled differently. The first type (red) indicates that the QTN locus genotype is a favorable allele, the second type (blue) indicates that the QTN locus genotype is an unfavorable allele, and the third type (black) indicates alleles with little or no influence. Unknown indicates that it was not detected, thus forming a calculable and traceable structured genetic knowledge graph, such as... Figure 3 As shown.

[0033] The genetic effect value can be calculated using the following formula. , where n represents the number of statistical trials in a single batch, and P represents the genetic effect value of a single batch (or prediction). This represents the measured value of a certain trait in a non-wild-type plant (mutant or heterozygous, etc.). This represents the measured value of wild-type plants for the same trait.

[0034] like:

[0035] ID 1, Gene `OsNPF6.1`, Chromosome `Chr1`, Agronomic Trait `Other`, Wild-type Phenotype `Low Nitrogen Fertilizer Use Efficiency`, Mutant Phenotype `Improved Nitrogen Fertilizer Use Efficiency`. Based on the efficiency improvement percentage mentioned in the report, a single-report normalized genetic effect value was set; for example, a 20% efficiency improvement would result in a genetic effect value of 0.2. For the same gene, multiple reported effect values ​​were recorded in the database for subsequent data training. That is, the genetic effect value was determined based on the ratio of the change in the mutant relative to the wild-type to the wild-type, or based on the ratio of the corresponding quantitative trait result value of the mutant to the result value of the wild-type.

[0036] ID 32, gene `sd1`, chromosome `Chr1`, agronomic trait `plant morphology`, wild-type phenotype `tall`, mutant phenotype `semi-dwarf`, genetic effect value is the percentage of reduced stem length.

[0037] ID 133, gene `Waxy`, chromosome `Chr6`, agronomic trait `taste quality`, wild-type phenotype `increased amylose content`, mutant phenotype `decreased amylose content`, genetic effect value is the proportion of increase or decrease in starch content.

[0038] Step 1.2: Sample Genotyping and Structured Phenotype Mapping

[0039] The system receives high-throughput sequencing data files from rice samples, uses the bioinformatics tool RiceNavi to perform genotyping at 320 predefined QTN loci, and converts the data into phenotypes according to predefined mapping rules.

[0040] Taking sample ZKW24001103 as an example, the generated genotypes can be as follows:

[0041] (1) 0|0: Wild type — corresponding to wild type phenotype.

[0042] (2) 1|1、DEL|-(number):——corresponds to mutant phenotype.

[0043] (3) 0|1: Heterozygous type — corresponding mutant phenotype.

[0044] (4) . / .、|:Unknown——The genotype of a certain gene was not detected, and the phenotype corresponding to the genotype is unknown.

[0045] (5) 0|2: Heterozygous alleles - the phenotype corresponding to the genotype is unknown.

[0046] (6) 1|2, 2|2: Multiple allele mutants - the phenotype corresponding to the genotype is unknown. Based on the above situations, the phenotypes corresponding to the 320 QTN genotypes of sample ZKW24001103 can be obtained.

[0047] Step 2: Based on the literature database, perform validation data retrieval for each QTN locus, and extract the planting location and environmental information contained in the validation data, as well as the phenotypic information under the disclosed planting environment.

[0048] Some literature data includes actual field planting results. These experimental results often provide clear data on quantitative traits with actual measurements. However, information on planting location and environment is often not presented in a systematic data format. If the literature contains specific conditions such as light, water, and soil, these can be used directly. If there are no direct planting conditions, but rather data based on geographical and date identifiers, these data need to be standardized. Taking environmental data as an example, several standardized conditions are set, including light conditions, temperature, and water conditions (including precipitation and the completeness of water conservancy projects; if there is no data on water conservancy projects, precipitation conditions are used). Based on geographical and date identifiers, the light, temperature, and water conditions of the target area are obtained from the China Meteorological Data Center. Preferably, each condition is normalized or graded based on the highest or median value. Step 3: For each QTN locus ID in the database, after removing outliers from the genetic effect values ​​of all experiments containing that locus, the standardized effect value of each QTN locus is obtained by averaging.

[0049] In a preferred implementation, step 3 includes: constructing a convolutional neural network model CNN, using the IDs of QTN loci, phenotypic information, and genetic effect values ​​from previous experiments in the database as training data, using the QTN loci and corresponding phenotypic information contained in the database as input, and using the standardized effect values ​​of the corresponding phenotypic information as output information to train the convolutional neural network model and obtain the standardized effect value of each QTN locus.

[0050] A Convolutional Neural Network (CNN) consists of an input layer containing the IDs and phenotypic information of QTN loci, convolutional layers, activation function layers, pooling layers, fully connected layers, and an output layer. The input layer takes each QTN locus as input, along with phenotypic information based on corresponding reports or data. It calculates the genetic effect value of each report using the quantitative trait results obtained from each report, and trains the model. After training, each QTN locus is fed back into the model to form the comprehensive genetic effect value corresponding to that locus.

[0051] Step 4: Construct an LSTM network based on fusion vectors. This fully connected network includes a gene information input branch and an environmental information input branch. The gene information input branch inputs the genotype information, phenotypic information, and standardized effect values ​​of QTN loci (especially mutant loci) contained in the validated varieties. The environmental information input branch inputs planting condition information. The planting condition information in the validation data, as well as the phenotypic information and standardized effect values ​​of quantitative traits disclosed in the planting environment, are used as data to train the fully connected network to obtain the trained LSTM network based on fusion vectors.

[0052] The LSTM network comprises multiple cascaded units, each unit including an input gate i(t), an output gate O(t), and a forget gate f(t), as follows: Figure 2 As shown.

[0053] The input gate i(t) uses the sigmoid function.

[0054] Potential update vectors are generated through the tanh layer. The two are combined to update the current cell state, where σ represents the sigmoid function, and h t-1 For the previous neuron, X t Input to the current neuron:

[0055] , where w i and w c Let b represent the weight matrix in the two equations. i and b c This represents the bias vector.

[0056] The unit output h of the previous time series t-1 The input vector X of the current time series is passed to the forget gate. t The output vector is f(t), and the formula for calculating f(t) is as follows: , where w f Represents network weights and b f This represents the bias vector.

[0057] By using the forget gate and the input gate, the previous time series state C can be realized. t-1 up to the current time series state C t The update, namely: ,

[0058] The formula for calculating the output gate is:

[0059] ,

[0060] Where wo Represents network weights and b o This represents the bias vector. The output gate outputs the predicted value of the target trait for the current variety and growing environment, or the predicted effect value of the target trait.

[0061] Specifically, for the gene information input branch, each parameter among the included gene ID, gene name, chromosome, associated trait, wild-type / mutant phenotype, and standard genetic effect value is used as a dimension of a vector to construct the first multidimensional vector X. t1 ={X 1， X 2， X 3，…. X S1 For situations where multiple traits need to be predicted, each trait is treated as a dimension, and each target trait is standardized. For example, salt tolerance can be divided into four different salt tolerance levels: none, low, medium, and high.

[0062] For the planting environment input branch, each parameter among temperature, light duration, rainfall during the growing season, rainfall during the maturity period, and fertilization status (after standardization) is used as a dimension of a vector to construct a second multi-dimensional vector, X. t2 ={X S1+1， X S1+2， X S1+3，…. X S1+S2 The second multidimensional vector is multiplied by a weighting coefficient, which represents the improvement in cultivation techniques, and then concatenated with the first multidimensional vector to obtain the concatenated input vector X. t ={ X 1， X 2， X 3，…. X S1, λ1(t)X S1+1， λ2(t)X S1+2， λ(t)X S1+3，…. λ S2 (t)X S1+S2 During the iterative optimization process, the value of λ(t) is adjusted synchronously, thereby incorporating the cultivation technology improvement factors that change over time into the prediction results. In the time series data, data from the same year are still arranged sequentially. [Xt] L , Xt N If they belong to the same year, then after training is completed, [Xt] will be... L , Xt NThe weights of the intervals are averaged to form a technical weighted value for the same year. It should be noted that different quantitative traits correspond to different weights. For example, seed coat color is often more affected by genes than by cultivation, while the seed setting rate is often more affected by cultivation. Therefore, training should be conducted separately for different quantitative traits.

[0063] The obtained spliced ​​vector is input into the input gate of the LSTM network. Different output gate prediction parameters are selected according to different prediction directions, and the networks are trained separately to form multiple parallel LSTM prediction models for predicting different traits.

[0064] Step 5: Extract the QTN loci and target planting environment information of the variety to be predicted, especially the mutant loci. Input the extracted QTN loci and target planting environment information of the variety to be predicted into the trained LSTM network based on fusion vector for prediction. Obtain the corresponding prediction results based on the input target trait.

[0065] 80% of the data containing actual planting information retrieved from the literature database was used as training data, and 20% of the data containing actual planting information was used as test data. The two models were tested using the test data. The first model used the improved LSTM model of this invention combined with time-series data input. The second model directly used a pre-built CNN model, processing the data (site data and environmental data, but without adding time-series attributes) according to the structured processing method of this invention, and then directly using it as vector input for training and testing. The test results are as follows:

[0066] Heading stage 3.5% (calculated based on an average heading period of 70 days) 4.7% Salt resistance 2.1% 2.9% 1000 grains weight 2.8% (based on an average thousand-grain weight of 25g) 5.2% Fruit setting rate 3.7% 6.9%

[0067] The comparison above shows that, in predicting the growth of rice itself (heading stage), the accuracy of the model of this invention is improved compared with the existing model, but the difference is not particularly significant. However, the accuracy of the prediction is significantly improved in terms of aspects related to cultivation techniques.

[0068] The test results show that the method of this invention can effectively predict the quantitative trait genes after rice planting.

[0069] The method of this invention can also be used for risk assessment:

[0070] For example, conduct systematic risk assessments on aspects such as resistance to biological stress, tendency to accumulate heavy metals, and key quality characteristics.

[0071] The scope of application of this invention is not limited to the examples above. This invention can be used for the prediction of major agronomic traits: outputting predicted heading date, plant height, rice quality characteristics, etc.; this invention can be used for material risk assessment: clearly indicating the risk level of the material in terms of rice blast, bacterial blight, cadmium accumulation, etc.

[0072] Of course, this invention is more effective in predicting quantitative traits, such as thousand-grain weight, seed setting rate, amylose content, gelatinization temperature, and typical characteristics (grain tip, seed coat color).

[0073] For the target variety, cultivation recommendations can be made to growers based on the predicted risks: specific field management recommendations can be given based on information such as tillering angle, lodging risk, and nitrogen fertilizer absorption efficiency; and the suitable planting areas, rice genus (indica / japonica), rice type (glutinous rice / soft rice), and type (conventional rice / hybrid rice) can be determined by combining factors such as resistance, rice quality, and heterozygosity.

[0074] Based on the evaluation of this invention, technicians can also provide targeted improvement suggestions: according to the genetic defects of the material (such as weak disease resistance and low nitrogen absorption efficiency), recommend the parents and target genes that can be introduced, and provide specific breeding improvement paths.

[0075] Although the principles of the present invention have been described in detail above with reference to preferred embodiments, those skilled in the art should understand that the above embodiments are merely illustrative explanations of the implementation of the present invention and are not intended to limit the scope of the present invention. The details in the embodiments do not constitute a limitation on the scope of the present invention. Any obvious changes, such as equivalent transformations or simple substitutions, based on the technical solutions of the present invention without departing from the spirit and scope of the present invention fall within the protection scope of the present invention.

Claims

1. A QTN data analysis method for intelligent navigation breeding based on a large model, characterized in that, The method includes: Step (1) Construct a structured QTN-phenotype-genetic effect value database, which includes the ID, phenotypic information, and genetic effect values ​​of each QTN locus in previous experiments; Step (2) Based on experimental data and literature databases, perform validation data retrieval for each QTN locus, extract the planting location and environmental information and the phenotypic information under the disclosed planting environment contained in the validation data, add time tags to all data, and sort them in chronological order to form a time data sequence; Step (3) For each QTN locus ID in the database, average the genetic effect values ​​of each experiment containing that locus to obtain the standardized effect value of each QTN locus; Step (4) Construct a fusion vector-based... The LSTM network based on fusion vectors receives gene information and environmental information respectively. The gene information includes gene information, phenotypic information and standardized effect value of QTN loci contained in the variety or gene to be predicted. The environmental information includes location and planting environment information. The planting location and environmental information contained in the time data sequence and the phenotypic information in the disclosed planting environment are used as data to train the network to obtain the trained LSTM network based on fusion vectors. Step (5): Extract the QTN loci of the variety to be predicted and the target planting environment information. Convert the extracted QTN loci contained in the variety to be predicted and the target planting environment information into fusion vectors and input them into the trained LSTM network based on fusion vectors for prediction.

2. The method for QTN data analysis in intelligent navigation breeding based on a large model according to claim 1, characterized in that, The initial genetic effect value is calculated according to the following formula. , where n represents the number of statistical trials in a single batch, and P represents the genetic effect value in a single batch. This represents the measured value of a mutant plant in a certain trait. This represents the measured value of wild-type plants for the same trait.

3. The QTN data analysis method for intelligent navigation breeding based on a large model according to claim 1, characterized in that, For the gene information input branch, each parameter is treated as a dimension of a vector to construct the first multidimensional vector X. t1 ={X 1， X 2， X 3，…. X S1 For the planting environment input branch, each planting environment-related parameter is used as one dimension of a vector to construct a second multi-dimensional vector, X. t2 ={X S1+1， X S1+1， X S1+1，…. X S1+S2 The second multidimensional vector is multiplied by a weighting coefficient, which represents the improvement in cultivation techniques, and then concatenated with the first multidimensional vector to obtain the concatenated input vector X. t ={ X 1， X 2， X 3，…. X S1, λ1(t)X S1+1， λ2(t)X S1+2， λ(t)X S1+3，…. λ S2 (t)X S1+S2 }, where S1 represents the dimension of the first multidimensional vector, S2 represents the dimension of the second multidimensional vector, λ1(t), λ2(t), ..., λ S2 (t) represents the weight of each dimension in the second multidimensional vector.

4. The QTN data analysis method for intelligent navigation breeding based on a large model according to claim 1, wherein the LSTM network includes an input gate i(t), an output gate O(t), and a forget gate f(t).

5. The QTN data analysis method for intelligent navigation breeding based on a large model according to claim 1, wherein the database contains at least 320 QTN loci.

6. The method for QTN data analysis in intelligent navigation breeding based on a large model according to claim 3, characterized in that, Related traits include one or more of the following: disease resistance, drought tolerance, salt tolerance, nitrogen uptake efficiency, lodging resistance, amylose content, gelatinization temperature, seed setting rate, seed coat color, thousand-grain weight, heading date, plant height, and rice quality characteristics.

7. The method for QTN data analysis in intelligent navigation breeding based on a large model according to claim 3, characterized in that, Genetic information includes parameters such as gene ID, gene name, chromosome, associated traits, wild-type / mutant phenotype, and standard genetic effect value.

8. The method for QTN data analysis in intelligent navigation breeding based on a large model according to claim 3, characterized in that, Planting environment parameters include one or more of the following: temperature, light duration, rainfall during the growing season, rainfall during the maturity period, and fertilization status.

9. The method for QTN data analysis in intelligent navigation breeding based on a large model according to claim 3, characterized in that, The step (3) further includes constructing a convolutional neural network model, using the QTN locus ID, phenotypic information and genetic effect values ​​of previous experiments in the database as training data, using the QTN loci and corresponding phenotypic information contained in the database as input, and using the standardized effect values ​​of the corresponding phenotypic information as output information to train the first convolutional neural network model and obtain the standardized effect value of each QTN locus.

10. A method for breeding using the analytical method according to any one of claims 1-9, characterized in that, The method includes analyzing and predicting the target quantitative traits according to the method before conducting breeding experiments.

Images