Multi-level evidence and artificial intelligence character key gene mining method and system

Abstract

The invention belongs to the technical field of gene mining, and particularly relates to a multi-level evidence and artificial intelligence character key gene mining method and system, and the method comprises the steps: S1, constructing a multi-omics database and integrating an interaction network; s2, predicting by using different machine learning algorithms; s3, constructing and applying an online platform: developing the online platform based on an integration result, and integrating a part of known interaction relationship databases; and systematically analyzing a genetic network of plant height, grains and drought resistance characters, and mining key genes. According to the method, the interaction network and multi-omics data are integrated, the mutual relation between functional genes is enriched, and all related genes of target characters are mined. And meanwhile, quantitative integration is performed on the relationship of the genes by utilizing artificial intelligence, and the importance of the related genes can be ranked. By means of the method, all related genes of the target character can be found out, key genes can be verified preferentially according to importance, and the method has important guiding significance on target character inheritance and molecular analysis.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of gene mining, but is not limited to it, and particularly relates to a multi-level evidence and artificial intelligence trait key gene mining method and system. BACKGROUND

[0002] With the development of high-throughput omics technology, modern biological research has expanded from single-dimensional data acquisition (such as genome, transcriptome) to multi-omics integrated analysis, including genomics, transcriptomics, proteomics, metabolomics, etc. These omics levels provide multi-dimensional information from DNA sequences to phenotype formation, providing a theoretical basis and technical support for systematic analysis of complex biological traits and identification of functional genes. The current traditional gene function mining method is usually time-consuming, labor-intensive, low-throughput, and difficult to effectively integrate multi-omics data and gene interaction relationships. The integrated map of maize gene expression and phosphorylation published by Walley et al. in 2016 shows that the addition of proteomic data greatly improves the quality of gene regulation networks based on transcriptome alone (Walley et al., 2016), and the first generation of multi-omics integrated network map provides a comprehensive network map of gene expression and regulation during maize development by integrating multiple omics data, which is of great significance for understanding the gene regulation mechanism of maize and guiding future maize improvement (Han et al., 2023). Multi-omics technology can elucidate the correlation mechanism between molecular regulation and phenotype, and systematically analyze the function and regulation mechanism of biomolecules. Moreover, multi-omics data resources can also be mutually verified, reducing false positives caused by single-omics analysis, improving research reliability, and obtaining more comprehensive and accurate regulation information through multi-omics joint analysis. In the face of massive multi-omics data, traditional statistical methods and candidate gene screening methods have obvious limitations in accuracy, scalability, and interpretability. The rapid development of machine learning (Machine Learning), especially deep learning (Deep Learning), has brought new opportunities for high-dimensional omics data modeling and analysis. Tao et al. used single-omics data to screen NETosis-related genes (NRG) through three machine learning methods, and explored their potential biological functions, immune cell invasion and clinical value (Tao et al., 2023).

[0003] Search results show that US patent WO2022197336A1 describes a method for discovering plant trait-related genes based on explainable machine learning and feature importance mechanisms. This method obtains gene expression profiles from plant tissue samples, inputs them into a prediction model to predict the target phenotype, and then uses explainable AI to analyze the model's decision-making mechanism to identify candidate genes that contribute the most to phenotype prediction. [1] This has a large overlap in technical ideas with your method of multi-omics data integration and machine learning to predict key trait genes, especially in the use of AI-driven feature importance to locate key genes.

[0004] Although this patent utilizes interpretable machine learning to screen key genes, it primarily relies on a single type of expression data input, lacking comprehensive utilization of features from multiple omics levels (such as metabolomics, proteomics, interaction networks, evolutionary attributes, etc.). Furthermore, the method focuses on key features output by the model, without fully considering the integration of gene interaction network-based scoring mechanisms between positive and candidate samples, cross-omics co-expression, traditional genetic association analyses such as GWAS / TWAS / eQTL, and the evolutionary selection features of the genes themselves (Pi, Tajima's D, dN / dS). This single-level dependence results in weak model generalization ability, potential omission of key genes, and limits the system's practicality and robustness in various trait and complex breeding applications. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method and system for mining key genes of traits using multi-level evidence and artificial intelligence.

[0006] This invention is implemented as follows: a method for mining key genes of traits based on multi-level evidence and artificial intelligence, comprising:

[0007] S1, construction of multi-omics databases and integration of interaction networks;

[0008] S2 uses different machine learning algorithms for prediction;

[0009] S3, Online Platform Construction and Application: Develop an online platform based on the integration results, integrate some known interaction relationship databases; systematically analyze the genetic networks of plant height, grain size, and drought resistance traits, and mine key genes.

[0010] Further, S1 also includes: collecting the existing interaction network of the target species and the published genes related to the target trait (defined as positive sample genes), and finding all possible candidate genes of the target trait according to the positive sample genes and the network. At the same time, the remaining genes in the whole genome except the positive sample genes and the candidate genes are named as random genes. The genes in the whole genome are scored by using the gene relationship contained in the interaction network. The specific score is based on the connection strength between the gene and all positive sample genes, and the score of one layer of direct connection is 1, and the score of two layers of connection is 0.01. According to the connection level with the known genes, the scores obtained by each gene are accumulated and summed to obtain the network score of the gene as a characteristic value.

[0011] Further, S1 also includes: integrating the related multi-omics data such as transcriptome, translation group and metabolome of the target trait. At the three omics levels, the co-expression coefficients between genes are calculated respectively, and then the relationship tightness with the known genes is measured. Taking co-transcription as an example, the expression amount of all genes in the whole genome is calculated, and the Pearson correlation coefficient between the expressions of different genes (only selecting the values with P less than 0.05) is calculated. The correlation coefficient of all genes with the known genes is screened, and the maximum value is selected as the representative of the relationship between the gene and the known gene as a characteristic value. Similarly, the characteristic values of co-translation and co-metabolism are obtained.

[0012] Further, S1 also includes: collecting the genotype data of natural population and the phenotype data of the target trait and performing outlier cleaning and arrangement. The correlation analysis between genes and phenotypes is performed. This includes GWAS analysis between genes and the target trait to identify genetic variations directly related to the trait; TWAS analysis between expression genes and the target trait to mine expression genes affecting the trait; and eQTL analysis between genes and the expression of known related genes. GWAS, TWAS and eQTL are all analyzed and calculated by using the existing related models. For the detection of epistasis, according to the two-by-two combination of all sites and the sites containing the known genes, according to the genotypes and population phenotypes, two-factor analysis of variance is performed, and the P value is used as the characteristic value of the site.

[0013] The characteristic analysis of all genes focuses on the attributes of the genes themselves, including the selection pressure indicators of the genes in the evolution process, such as Pi value, dN / dS ratio and Tajima's D value. These attributes reflect the genetic variation and evolutionary history of the genes in the population, which are crucial for understanding the function and adaptability of the genes.

[0014] Further, S2 further comprises: integrating all of the feature values to obtain a matrix of all genomic genes and all omics feature values to obtain 13 important features, whole transcriptome association (TWAS-P), core co-expression of genes at three levels (Corexp1, Corpro1, Coribo1), known genes downstream of genes at different omics levels (Rna-eqtl-up, Ribo-eqtl-up, Pro-eqtl-up, Rna-eqtl-down, Ribo-eqtl-down, Pro-eqtl-down), network features (Netscore), maximum value in different levels (Cor-all, eqtl-up-all, eqtl-down-all), whole genome association analysis (GWAS-P), and maximum expression amount in the whole growth period (MAX).

[0015] Further, S2 further comprises: using as many machine learning methods as possible, such as Bayesian, support vector machine, random forest, decision tree, and gradient boosting tree (GBT) model, to classify the samples. Since a single model is more sensitive to noise or outliers in data when classifying, it cannot fully utilize the information in the data, resulting in poor prediction performance. It is known that stacking ensemble methods are used for genomic prediction to improve prediction accuracy. There are several ensemble learning methods, such as Bagging, Boosting, and Stacking, among which the Stacking method improves prediction accuracy by integrating the outputs of different models into a meta-model. Therefore, the above single models are stacked into an ensemble, a Logistic model is used as a meta-model, and the outputs of the single models are stacked as new inputs to fit the Logistic model to compensate for the shortcomings of the single models and improve prediction performance. In order to solve the overfitting problem of the ensemble method, the present application improves the above-mentioned ensemble method by introducing a K-fold cross-validation method. The output results of the single models are 5-fold cross-validated, K-1 folds are trained each time, and the prediction value of the remaining 1 fold is used as the input of the meta-model to effectively prevent overfitting. By integrating the advantages of different models, the overall prediction performance and the generalization ability to new data can be improved, while the risk of overfitting can be reduced, thereby better adapting to different data scenarios and achieving better classification results. The best machine learning model is selected for the new algorithm.

[0016] Another object of the present application is to provide a multi-level evidence and artificial intelligence trait key gene mining system for realizing the multi-level evidence and artificial intelligence trait key gene mining method, comprising:

[0017] The database construction module is used for constructing a multi-omics database and integrating an interaction network.

[0018] A machine learning prediction module for making predictions using different machine learning algorithms;

[0019] An online platform building and application module for developing an online platform based on the integration results, integrating part of the known interaction relationship database; the system analyzes the genetic network of plant height, grain and drought resistance traits, and mines key genes.

[0020] Another object of the present application is to provide a computer device comprising a memory and a processor, the memory storing a computer program, the computer program being executed by the processor to make the processor execute the steps of the multi-level evidence and artificial intelligence trait key gene mining method.

[0021] Another object of the present application is to provide a computer readable storage medium storing a computer program, the computer program being executed by the processor to make the processor execute the steps of the multi-level evidence and artificial intelligence trait key gene mining method.

[0022] Another object of the present application is to provide an information data processing terminal comprising the multi-level evidence and artificial intelligence trait key gene mining system.

[0023] In combination with the above technical solutions and the technical problems solved, the technical solution to be protected by the present application has the following advantages and positive effects:

[0024] First, the present application solves the one-sidedness of single omics analysis by analyzing multiple omics data and constructing interaction networks of trait-related genes, further increasing the understanding of trait genetic structure and biological mechanisms. The present application integrates interaction networks and multi-omics data, enriches the mutual relationship between functional genes, and mines all related genes of the target trait. At the same time, the use of artificial intelligence to quantitatively integrate the relationship between genes can rank the importance of related genes. Through the present application, not only can all related genes of the target trait be found, but also key genes can be verified according to importance, which has important guiding significance for the genetic and molecular analysis of the target trait.

[0025] Second, the technical solution of the present application solves the technical problem that people have been eager to solve but have failed to succeed:

[0026] With the development of sequencing technology, a large amount of various omics data of organisms has emerged, but the integration method between omics data is still less. We first analyze the data within each omics level, and then use artificial intelligence to integrate and analyze between omics, fully utilize big data to mine related signals, and make the gene mining of the target trait more accurate and comprehensive.

[0027] The technical scheme of the present application overcomes the technical bias: genomic breeding is the current mainstream design breeding form, which is based on whole genome or as many molecular markers as possible for parameter evaluation and phenotype prediction, which improves the efficiency of breeding to a certain extent. However, genomic breeding faces the statistical barrier of "big P, small N", and the biological "burden" of undesirable linkage of global molecular markers, and we urgently need to conduct global functional gene mining and molecular mechanism analysis of complex agronomic traits, so as to achieve precise and intelligent design breeding. The classical functional gene mining of target traits is to use QTL mapping and map-based cloning of target phenotypes, which requires a lot of manpower and material resources, and often can only analyze the function of a single gene. With the development of biological big data, there are many methods at present to mine genes related to target traits at a single omics level and construct gene association networks. The information contained in such networks is often relatively single. The present research integrates interaction networks and multi-omics data, enriches the mutual relationship between functional genes, and mines all genes related to target traits. At the same time, the use of artificial intelligence to quantitatively integrate the relationship between genes can rank the importance of related genes. Through this method, not only can all genes related to target traits be found, but also key genes can be verified according to importance, which has important guiding significance for genetic and molecular analysis of target traits. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is a multi-level evidence and artificial intelligence trait key gene mining method flowchart provided by an embodiment of the present application;

[0029] Figure 2 is a multi-level evidence and artificial intelligence trait key gene mining system structure diagram provided by an embodiment of the present application;

[0030] Figure 3 is a first-generation maize interaction group network (Han et al., 2023, Nature Genetics) provided by an embodiment of the present application;

[0031] Figure 4 is a distribution diagram of different categories of genes in different omics provided by an embodiment of the present application;

[0032] Figure 5 is a diagram of characteristic values of different categories of genes provided by an embodiment of the present application;

[0033] Figure 6 is a GWAS and TWAS analysis result diagram of kernel hundred-grain weight provided by an embodiment of the present application;

[0034] Figure 7 is a prediction effect diagram provided by an embodiment of the present application. DETAILED DESCRIPTION

[0035] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present application and not to limit the present application.

[0036] The prior art generally relies on single-omics level association analysis or limited interaction network information in key trait gene mining, which often makes it difficult to fully reveal the genetic basis of complex traits. For example, although traditional GWAS can locate genetic variations significantly related to phenotypes, its resolution capability is often limited by factors such as population structure, linkage disequilibrium and statistical threshold, resulting in some key gene signals being masked. At the same time, relying solely on transcriptome co-expression or protein interaction network information is also prone to problems such as large noise interference and insufficient generalization ability, lacks comprehensive support from cross-level evidence, and is difficult to form a stable and reliable candidate gene set.

[0037] The present method integrates multi-level evidence, which theoretically breaks through the limitations of the prior art. The working principle is first to deeply integrate multi-omics data and interaction networks, and to quantify the network connection relationship of positive sample genes and candidate genes, giving each gene a clear network characteristic value. This layered weighted network scoring method avoids the roughness of single distance threshold screening, enabling potential candidate genes to obtain continuous quantitative indicators based on their connectivity with known genes, improving the accuracy and explanatory power of feature representation.

[0038] On the multi-omics level, the method further integrates transcriptome, translation and metabolome information, and selects the co-expression characteristics of genes and known genes at different levels by calculating the Pearson correlation coefficient and strictly limiting the significance threshold. This process not only enhances the ability to describe the functional coupling relationship between genes, but also makes up for the shortcomings of single-omics analysis in dynamic change and upstream and downstream regulation analysis. Taking the maximum correlation coefficient as the representative characteristic value helps to stably capture the most significant molecular level association, thereby improving the robustness of the prediction model.

[0039] Another core step of the method is the systematic analysis of genotype and phenotype data of natural populations. GWAS is used to identify genetic variations directly related to traits, TWAS is used to find genes closely related to trait expression levels, and eQTL is used to reveal key links in regulatory relationships. The joint application of the three forms a complementary framework to ensure that both direct effects and indirect regulatory effects can be captured. In addition, the introduction of epistasis detection enables the interaction between loci to be quantified as a characteristic value, expanding the adaptability of the model to complex genetic backgrounds.

[0040] The evolutionary properties of the genes themselves are also highlighted in the method. By calculating the Pi value, dN / dS ratio and Tajima's D value, the method can reflect the genetic variation pattern and selection pressure of the gene at the population level. These indicators not only increase the biological interpretability of the model, but also provide evolutionary level evidence to support the functional positioning of key genes. Compared with traditional methods, this step injects population genetics depth information into the model, making it more consistent with the real background of the genetic mechanism of the trait.

[0041] All features are integrated into a genome-wide feature matrix, covering thirteen core indicators, including whole transcriptome correlation, three-layer co-expression value, upstream and downstream regulatory gene information, network features, maximum expression and its standard deviation during the whole growth period. By inputting different machine learning algorithms into this matrix for modeling and training, the cross-level evidence can be maximally utilized to improve the prediction accuracy and generalization ability. The experimental results show that this method is significantly better than traditional single methods in the key gene mining of plant height, grain and drought resistance traits, and has the potential for wide application in molecular breeding and functional gene analysis.

[0042] As shown in Figure 1 , the trait key gene mining algorithm based on multi-level evidence includes the following steps:

[0043] (1) Multi-omics database construction and integration of interaction networks

[0044] ① Collect the existing interaction networks of the target species and the published genes related to the target trait (defined as positive sample genes). According to the positive sample genes and the network, all possible candidate genes of the target trait are found. At the same time, the remaining genes in the whole genome except the positive sample genes and the candidate genes are named as random genes. The genes in the whole genome are scored using the gene relationships contained in the interaction network. The specific score is based on the connection strength between the gene and all positive sample genes, with a score of 1 for one layer of direct connection and a score of 0.01 for two layers of connection. According to the connection level with known genes, the obtained score of each gene is summed to obtain the network score of the gene as a feature value.

[0045] ② Integrate the multi-omics data such as transcriptome, translation and metabolome related to the target trait. At these three omics levels, the co-expression coefficients between genes are calculated to measure the closeness of the relationship with known genes. Taking co-transcription as an example, the expression of all genes in the whole genome is calculated, and the Pearson correlation coefficient between different gene expressions is calculated (only selecting values with P less than 0.05). The correlation coefficient between all genes and known genes is selected, and the maximum value is selected as the representative of the relationship between the gene and the known gene as a feature value. Similarly, the co-translation and co-metabolism feature values are obtained.

[0046] ③ Collect genotypic data of natural populations and phenotypic data of the target trait, and perform outlier cleaning and rectification. Perform gene-phenotype association analysis. This includes GWAS analysis between genes and the target trait to identify genetic variations directly related to the trait; TWAS analysis between expressed genes and the target trait to identify expressed genes influencing the trait; and eQTL analysis between genes and the expression of known related genes. GWAS, TWAS, and eQTL are all performed using existing correlation models. For epistasis detection, a two-way ANOVA is conducted based on pairwise combinations of all loci with known gene-containing loci, according to their genotype and population phenotype, using the p-value as the locus's eigenvalue.

[0047] ④ The characterization of all genes focuses on their intrinsic properties, including indicators of selection pressures during evolution, such as Pi value, dN / dS ratio, and Tajima's D value. These properties reflect the genetic variation and evolutionary history of genes within a population and are crucial for understanding gene function and adaptation.

[0048] (2) Using different machine learning algorithms for prediction

[0049] ① Integrating all feature values ​​yields a matrix of whole-genome genes and all omics feature values, resulting in 13 important features: Transcriptome-wide association score (TWAS-P), co-expression of genes at three levels (Corexp1, Corpo1, Coribo1), upstream and downstream regulatory genes of known genes at different omics levels (Rna-eqtl-up, Ribo-eqtl-up, Pro-eqtl-up, Rna-eqtl-down, Ribo-eqtl-down, Pro-eqtl-down), network features (Netscore), maximum values ​​at different levels (Cor-all, eqtl-up-all, eqtl-down-all), genome-wide association analysis (GWAS-P), and maximum expression level throughout the reproductive period (MAX).

[0050] ② Use as many machine learning methods as possible, such as logistic regression, Bayesian, support vector machine, random forest, decision tree and gradient boosting tree (GBT) model to classify samples. Since a single model is sensitive to noise or outliers in the data when classifying, it cannot fully utilize the information in the data, resulting in poor prediction performance. It is known that stacking ensemble methods are used for genomic prediction to improve prediction accuracy. There are several methods of ensemble learning, such as Bagging, Boosting and Stacking, among which the Stacking method improves prediction accuracy by integrating the outputs of different models into a meta-model. Therefore, the above single models are stacked, the Logistic model is used as the meta-model, and the outputs of the single models are stacked as new inputs to fit the Logistic model to compensate for the shortcomings of the single models and improve the prediction performance. In order to solve the overfitting problem of the ensemble method, the present application improves the above ensemble method by introducing the K-fold cross-validation method. The output results of the single models are 5-fold cross-validated, K-1 folds are trained each time, and the prediction value of the remaining 1 fold is used as the input of the meta-model, which effectively prevents overfitting. By integrating the advantages of different models, the overall prediction performance and the generalization ability to new data can be improved, while the risk of overfitting can be reduced, so that different data scenarios can be better adapted and better classification results can be achieved. Select the best machine learning model for the new algorithm.

[0051] (3) Online platform building and application

[0052] ① Based on the integrated results, an online platform is developed, and part of the known interaction relationship database is integrated.

[0053] ② Taking corn as an example, the genetic network of its plant height, grain, drought resistance and other traits is analyzed, and key genes are mined.

[0054] As shown in Figure 2 , the multi-level evidence and artificial intelligence trait key gene mining system comprises:

[0055] A database construction module is used to construct multi-omics databases and integrate interaction networks.

[0056] A machine learning prediction module is used to predict using different machine learning algorithms.

[0057] An online platform building and application module is used to develop an online platform based on the integrated results, integrate part of the known interaction relationship database, analyze the genetic network of plant height, grain and drought resistance traits, and mine key genes.

[0058] The present application shows the analysis results of corn grain related trait key gene mining using multi-level evidence, and the specific process is shown in the following examples.

[0059] Example

[0060] 1. Integrating network features

[0061] Taking maize kernel traits as an example, we first collected 63 genes related to kernel development by searching and reading a large number of Chinese and English literature. We then located the annotations of these genes in the maize B73 V4 version reference genome, standardized their IDs and location information, and named them positive sample genes. The team previously constructed the first-generation interaction network for maize B73 (…). Figure 3 This dataset contains information on the transcriptome, translatome, proteome, and three-dimensional interactionome throughout the entire growth period (Han et al., 2023). Based on the known genes and interaction networks, a sub-network of genes related to grain traits was constructed. Genes outside the positive samples in the sub-network were named candidate genes, and the remaining genes across the entire genome after removing positive sample genes and candidate genes were named random genes. The genes across the entire genome were scored using the gene relationships contained in the interaction network. The specific score was based on the connection strength between the gene and all positive sample genes, with a score of 1 for a single-layer direct connection and 0.01 for a two-layer connection. The 63 scores obtained for each gene were summed based on the connection levels with 63 known genes to obtain the network score of that gene as a feature value.

[0062] 2. Integration of genes with known genes related to traits

[0063] Transcriptome, translatome, and metabolome data were collected from nearly 200 samples of seeds 6 days after pollination. This data was used to establish gene interactions at the transcriptional, translational, and metabolic levels during seed development. Co-expression, co-translation, and co-metabolism associations among genes were calculated at multiple omics levels. At each of these three omics levels, co-expression coefficients were calculated to measure the strength of relationships with known genes. Taking co-transcription as an example, the expression levels of all genes across the entire genome were calculated, and Pearson correlation coefficients between different gene expressions were calculated (only values ​​with p < 0.05 were selected). The correlation coefficients between all genes and 63 known genes were screened, and the highest value was selected as the representative of the relationship between this gene and known genes, serving as the eigenvalue. Similarly, eigenvalues ​​for co-translation and co-metabolism were obtained. Figure 4 ).

[0064] 3. Association analysis between genes and phenotypes

[0065] Phenotypes of 100-kernel weight of 508 maize inbreds were collected, and 453 materials with both genotypes and phenotypes were obtained. GWAS analysis between genes and target traits was performed to identify genetic variations directly related to traits; TWAS analysis between expressed genes and target traits was performed to mine expressed genes affecting traits; and eQTL analysis between genes and known related gene expression was performed. GWAS, TWAS and eQTL were all analyzed and calculated using existing correlation models. For the detection of epistasis, according to the two-way combination of all sites and known gene containing sites, according to their genotypes and population phenotypes, two-factor analysis of variance was performed, and P value was used as the eigenvalue of the site.

[0066] 4. Properties of genes themselves

[0067] Currently, a "super pan-genome" covering most materials in Zea mays has been published in corn, including 507 modern corn materials, 31 corn landrace materials and 183 corn wild relative switchgrass materials, covering all subspecies of Zea mays and having wide representativeness (Gui et al., 2022). Using this corn pan-genome data, relevant feature analysis of whole genome genes can be performed. The gene age of the whole genome gene, the Pi value, the dN / dS value and the Tajima D value of the gene in the process of selection and evolution, the number of tissues expressed by the gene in the whole growth period and the maximum expression amount are calculated as eigenvalues. Figure 5 ).

[0068] 5. Integrate the feature value matrix, mark positive and negative samples, and divide the training set

[0069] The characteristic values of the whole genome genes are integrated to obtain characteristics with significant differences, TWAS-P, Corexp1, Corpro1, Coribo1, Rna-eqtl-up, Ribo-eqtl-up, Pro-eqtl-up, Rna-eqtl-down, Ribo-eqtl-down, Pro-eqtl-down, Netscore, Cor-all, eqtl-up-all, eqtl-down-all, GWAS-P, and MAX. The published target trait-related gene markers are positive samples 1, and all possible candidate genes of the target trait are found according to the positive sample genes and the network. At the same time, the remaining genes in the whole genome range except the positive sample genes and the candidate genes are named as random gene markers as negative samples 0. 10% of the positive sample set factors and the same number of negative sample genes are fixed as the test set, 80% of the positive sample set factors and 10 times the number of negative sample genes are selected as the training set, each model is cycled 20 times, the remaining 10% of the positive sample set factors and 5 times the number of negative sample genes are used as the test set to evaluate the effect of each cycle, the optimal parameter combination is selected for the evaluation of the test set, and then the whole process is cycled 10 times to improve the accuracy of the model.

[0070] 6. Constructing a machine learning model to predict grain-related candidate genes

[0071] The characteristic value matrix of the whole genome genes is obtained, and support vector machines, decision trees, random forests, Bayesian, and gradient boosting trees (GBT) models are used for sample classification. Since a single model is more sensitive to noise or outliers in data when classifying, it cannot fully utilize the information in the data, resulting in poor prediction performance. There are studies that use stacked ensemble methods for genomic prediction to improve prediction accuracy. There are several ensemble learning methods, including Bagging, Boosting, and Stacking, among others. The stacking ensemble method improves prediction accuracy by integrating the outputs of different models into a meta-model. Therefore, the above single models are stacked into an ensemble, a Logistic model is used as a meta-model, and the outputs of the single models are stacked as new inputs to fit the Logistic model to compensate for the shortcomings of the single models and improve prediction performance.

[0072] To solve the overfitting problem of the integrated method, the application improves the above integrated method, introduces the K-fold cross-validation method, and performs 5-fold cross-validation on the output results of a single model. Each time, K-1 folds are trained, and the prediction value of the remaining 1 fold is taken as the input of the meta model, effectively preventing overfitting. By integrating the advantages of different models, the overall prediction performance and generalization ability to new data can be improved, while the risk of overfitting can be reduced, so as to better adapt to different data scenarios and achieve better classification results. The empirical results of 10 cycles of each model show that all the evaluated models have achieved good prediction performance. The average can reach more than 0.75. Figure 7 )

[0073] In the application, in order to overcome the overfitting problem of the integrated learning method in complex data scenarios, the research team introduces the K-fold cross-validation mechanism. Specifically, the output results of a single model are taken as the basis, and through 5-fold cross-validation, K-1 folds of samples are used for training each time, and the prediction value of the remaining 1 fold is input into the meta model, so that the model's dependence on the training set is broken during the training process, avoiding the risk of the model memorizing the training data. This improvement not only enhances the robustness of the model, but also shows better adaptability under different data sets and feature dimensions.

[0074] Through multi-model integration, the application makes full use of the complementary advantages of different algorithms. Support vector machine (SVM) is relatively weak in recall rate, but other models perform outstandingly in overall indicators, especially logistic regression, decision tree and random forest, which have reached a high level in accuracy, recall rate, F1 value and AUC comprehensive evaluation indicators. Through this integration strategy, the overall prediction performance of the model is improved, and its generalization ability to new data is significantly enhanced, effectively dealing with different types and sizes of data challenges.

[0075] It is worth noting that the empirical results show that in the recall rate, a key indicator, SVM performs the worst, only about 0.75, while the improved integrated method makes the precision of other models generally reach about 0.83. This result shows that the improved method can reduce the risk of "missed report", which is of great significance for biological research or medical diagnosis that needs to capture as many target samples as possible.

[0076] The improved integrated method of the application introduced K-fold cross-validation effectively avoids overfitting while ensuring high precision, improving the robustness of the model in complex scenarios. Combined with the results shown in Figure 7 It can be seen intuitively that the differences in prediction performance of different models and the improvement advantages. Compared with Figure 6GWAS and TWAS analysis results echo each other, and the method provides more reliable and efficient tool support for genetic mechanism analysis and gene prediction of complex traits.

[0077] It should be noted that the embodiments of the present application can be realized by hardware, software or a combination of hardware and software. The hardware part can be realized by special logic; the software part can be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or a specially designed hardware. Those skilled in the art can understand that the above-mentioned devices and methods can be realized by computer executable instructions and / or included in processor control code, for example, such code is provided on a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The device and its modules of the present application can be realized by hardware circuit such as ultra-large scale integrated circuit or gate array, semiconductor such as logic chip, transistor, or programmable hardware device such as field programmable gate array, programmable logic device, etc., can also be realized by software executed by various types of processors, and can also be realized by a combination of the above-mentioned hardware circuit and software, such as firmware.

[0078] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any modification, equivalent replacement and improvement within the technical range disclosed by the present application, which is within the spirit and principle of the present application, should be covered within the protection scope of the present application.

Claims

1. A method for mining key trait genes using multi-level evidence and artificial intelligence, characterized in that, include: S1 multi-omics database construction and interaction network integration; S2 uses different machine learning algorithms for prediction; S3, Online Platform Construction and Application: Develop an online platform based on the integration results, integrate some known interaction relationship databases; systematically analyze the genetic networks of plant height, grain size, and drought resistance traits, and mine key genes.

2. The method for mining key trait genes using multi-level evidence and artificial intelligence as described in claim 1, characterized in that, S1 also includes: collecting existing interaction networks and published genes related to the target trait in the target species; identifying all possible candidate genes for the target trait based on positive sample genes and the network; naming the remaining genes in the whole genome after removing positive sample genes and candidate genes as random genes; scoring the genes in the whole genome using the gene relationships contained in the interaction network; the specific score is based on the connection strength between the gene and all positive sample genes, with a score of 1 for a single-layer direct connection and a score of 0.01 for a two-layer connection; summing the scores obtained for each gene based on the connection level with known genes to obtain the network score of the gene as a feature value.

3. The method for mining key trait genes using multi-level evidence and artificial intelligence as described in claim 1, characterized in that, S1 also includes: integrating multi-omics data such as transcriptomics, translatomics, and metabolomics related to the target trait; calculating co-expression coefficients between genes at these three omics levels to measure the closeness of the relationship with known genes; calculating the expression levels of all genes in the whole genome, calculating the Pearson correlation coefficients between different gene expressions, and only selecting values ​​with P less than 0.05; screening out the correlation coefficients between all genes and known genes, and selecting the maximum value as the representative of the relationship between this gene and known genes, as the feature value; similarly, obtaining the feature values ​​of co-translation and co-metabolism.

4. The method for mining key trait genes using multi-level evidence and artificial intelligence as described in claim 1, characterized in that, S1 also includes: collecting genotype data of natural populations and phenotypic data of target traits and cleaning and organizing outliers; performing gene-phenotype association analysis; this includes GWAS analysis between genes and target traits to identify genetic variations directly related to traits; TWAS analysis between expressed genes and target traits to identify expressed genes that influence traits; and eQTL analysis between genes and the expression of known related genes; GWAS, TWAS, and eQTL are all performed using existing correlation models; for epistasis detection, a two-way ANOVA is performed based on pairwise combinations of all loci with known gene-containing loci, according to their genotype and population phenotype, with the p-value as the characteristic value of the locus; The characterization of all genes focuses on the properties of the genes themselves, including indicators of the selective pressures they have experienced during evolution, such as Pi value, dN / dS ratio, and Tajima's D value. These properties reflect the genetic variation and evolutionary history of genes in a population.

5. The method for mining key trait genes using multi-level evidence and artificial intelligence as described in claim 1, characterized in that, S2 also includes: integrating all feature values ​​to obtain a matrix of whole genome genes and all omics feature values ​​to obtain 13 important features, whole transcriptome association analysis, gene co-expression at three levels, upstream and downstream regulatory genes of known genes at different omics levels, network features, taking the maximum value at different levels, whole genome association analysis, and maximum expression level throughout the reproductive period.

6. The method for mining key trait genes using multi-level evidence and artificial intelligence as described in claim 1, characterized in that, S2 also includes: using as many machine learning methods as possible, including Bayesian, support vector machine, random forest, decision tree, and gradient ascending tree models to classify samples; since a single model is sensitive to noise or outliers in the data when classifying, it cannot fully utilize the information in the data, resulting in poor prediction performance; it is known that stacked ensemble methods have been used for genome prediction to improve prediction accuracy; there are several ensemble learning methods such as Bagging, Boosting, and Stacking, among which stacked ensemble (Stacking) improves prediction accuracy by integrating the outputs of different models into a meta-model; therefore, the above single models are stacked and integrated, using the Logistic model as the meta-model, and the outputs of the single models are stacked column by column as new inputs and fitted to the Logistic model to make up for the defects of the single models and improve prediction performance; in order to solve the overfitting problem of ensemble methods, the K-fold cross-validation method is introduced, and the output results of the single models are cross-validated in 5 folds, with K-1 folds trained each time, and the prediction value of the remaining 1 fold is used as the input of the meta-model.

7. A system for mining key trait genes using multi-level evidence and artificial intelligence as described in any one of claims 1 to 6, characterized in that, include: The database construction module is used for building multi-omics databases and integrating interaction networks; The machine learning prediction module is used to make predictions using different machine learning algorithms; The online platform construction and application module is used to develop an online platform based on the integration results, integrating some known databases of interaction relationships. The genetic networks of plant height, grain size, and drought resistance traits were systematically analyzed to identify key genes.

8. A computer device comprising a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the multi-level evidence and artificial intelligence trait key gene mining method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the multi-level evidence and artificial intelligence trait key gene mining method as described in any one of claims 1 to 6.

10. An information data processing terminal, comprising the multi-level evidence and artificial intelligence trait key gene mining system as described in claim 7.

Images