Biomarker combination and method for screening mungbean drought land cultivation adaptability

Abstract

This invention discloses a biomarker combination and method for screening the adaptability of mung beans to dryland cultivation, belonging to the field of breeding technology. The biomarker combination consists of ten metabolites: L-tyrosine, L-phenylalanine, apigenin, L-glutamic acid, D-galactose, tryptophan, 5-hydroxytryptophan, guanosine, L-arginine, and L-proline. The screening method involves obtaining the content data of the ten metabolites in the mung bean sample to be tested, standardizing the content using the mean content of irrigated mung bean populations as a benchmark, calculating normalized weights based on variable importance using a random forest classification model, and constructing a comprehensive score by weighted summation of the standardized relative contents. If the comprehensive score is not lower than a preset threshold, the mung bean sample to be tested is determined to have breeding value for dryland cultivation quality. This invention is mainly used for early screening of mung bean drought-resistant quality breeding.

Description

Technical Field

[0001] This invention relates to the field of breeding technology. More specifically, this invention relates to a combination of biomarkers and a method for screening the adaptability of mung beans to dryland cultivation. Background Technology

[0002] Screening mung bean germplasm with superior quality breeding value under dryland cultivation conditions is a crucial step in the breeding of drought-resistant mung bean varieties. In current breeding practices, the identification of the quality breeding value of mung beans cultivated in dryland mainly relies on field phenotypic observation and yield trait statistics. The identification cycle is long and greatly affected by fluctuations in environmental factors such as year and region, making it difficult to meet the needs of large-scale early screening of breeding materials.

[0003] Using metabolites as predictive indicators of plant phenotypes has become a research hotspot in recent years. Metabolites are the final products of the interaction between gene expression and environmental factors, and can directly reflect the physiological state of plants. Chinese invention patent application CN116381085A discloses a method for tracing the origin of mung beans based on characteristic differential metabolites. This method uses ultra-high performance liquid chromatography-time-of-flight mass spectrometry (UHPLC-TOF-MS) with non-targeted metabolomics to detect metabolite data of mung bean samples from different origins, uses chemometrics to screen characteristic differential metabolites, and constructs an origin tracing classification model through machine learning to achieve accurate origin tracing of mung beans. This scheme uses origin category as the classification target, and the screened differential metabolites are used to distinguish geographical origins. It does not address the impact of water supply conditions on mung bean quality, nor does it provide metabolite combinations and their application methods for evaluating the breeding value of mung beans grown in dryland conditions. In other words, existing crop classification studies based on metabolites are mostly aimed at origin tracing, lacking metabolite combinations that can be directly used for early screening in breeding practice, with the breeding value of mung beans grown in dryland conditions as the judgment target. Summary of the Invention

[0004] The purpose of this invention is to provide a combination of biomarkers and a method for screening the adaptability of mung beans to dryland cultivation, so as to at least solve the above-mentioned problems.

[0005] This invention provides a biomarker combination for screening the adaptability of mung beans to dryland cultivation, which consists of the following ten metabolites: L-tyrosine, L-phenylalanine, apigenin, L-glutamic acid, D-galactose, tryptophan, 5-hydroxytryptophan, guanosine, L-arginine, and L-proline; the dryland cultivation adaptability of mung beans is synergistically characterized by the difference in the relative content of the ten metabolites in dryland-cultivated mung beans and irrigated mung beans.

[0006] This invention also provides a method for early screening of the adaptability of mung beans for dryland cultivation, comprising the following steps: obtaining the content data of ten metabolites in the mung bean sample to be tested, the ten metabolites being L-tyrosine, L-phenylalanine, apigenin, L-glutamic acid, D-galactose, tryptophan, 5-hydroxytryptophan, guanosine, L-arginine, and L-proline; using the mean content of the corresponding metabolites in the irrigated mung bean population as a benchmark, performing standardization processing on the content data of the ten metabolites to obtain the standardized relative content of each metabolite; and based on the content of each metabolite in the dryland cultivation of mung beans... In the random forest classification model, the importance of variables is calculated using normalized weights. These normalized weights are then used to perform a weighted summation on the standardized relative content to construct a comprehensive score for the mung bean sample to be tested. When the comprehensive score is not lower than a preset threshold, the mung bean sample is determined to have breeding value for quality mung bean grown in dryland and is suitable for dryland cultivation. The preset threshold is determined based on the 10th percentile of the comprehensive score distribution of known dryland mung bean samples. Samples below the 10th percentile are considered suitable for irrigated cultivation, while samples above the 10th percentile are considered suitable for dryland cultivation.

[0007] Preferably, the method for constructing a random forest classification model includes the following steps: collecting mung bean samples from multiple known cultivation environments, measuring the content of ten metabolites—L-tyrosine, L-phenylalanine, apigenin, L-glutamic acid, D-galactose, tryptophan, 5-hydroxytryptophan, guanosine, L-arginine, and L-proline—in each sample, and using the content of these ten metabolites to construct a feature vector; constructing a training dataset using the cultivation environment category of the samples as a label; training a random forest classification model using the training dataset; evaluating the classification performance of the random forest classification model, and using a random forest classification model whose classification performance meets the preset requirements as a random forest prediction model for quality breeding of mung beans grown in dryland areas, and performing quality breeding screening for unknown mung bean samples grown in dryland areas.

[0008] Preferably, the standardization process uses the mean content of the corresponding metabolite in the irrigated mung bean population as the benchmark, calculates the ratio between the content of each metabolite in the mung bean sample to be tested and the corresponding mean, and then uses the logarithmic transformation result of the ratio to base 10 as the standardized relative content.

[0009] Preferably, before training the random forest classification model, a grid search is used to optimize the hyperparameters of the random forest classification model, including the number of decision trees, the maximum depth, and the minimum number of samples required for splitting. The average accuracy of cross-validation is used as the basis for hyperparameter selection, and the optimized hyperparameters are used to train the random forest classification model.

[0010] Preferably, stratified cross-validation is performed on the training dataset, dividing the samples in the training dataset into stratified groups according to the cultivation environment category, so that the class ratio of irrigated cultivation samples to dryland cultivation samples in the training and validation folds of each cross-validation fold is consistent with the overall class ratio of the training dataset; a random forest classification model is trained on each fold and the classification performance is evaluated on the corresponding validation fold, and the average value of the evaluation results of each fold is used as the cross-validation performance of the random forest classification model; the classification performance of the random forest classification model is double-verified based on the cross-validation performance and the evaluation results of the test set. When both satisfy the requirements of classification accuracy not less than 90%, sensitivity not less than 85%, and specificity not less than 85%, the random forest classification model is confirmed to have passed the performance verification.

[0011] Preferably, after constructing the training dataset, the number of irrigated cultivation samples and the number of dryland cultivation samples in the training dataset are counted, and the ratio of the number of samples of the two classes is calculated. When the ratio is greater than 2:1 or less than 1:1, class balancing is performed on the training dataset. The class balancing process uses synthetic minority class oversampling technology to generate synthetic samples for the class with fewer samples, so that the ratio of the number of samples of the two classes falls within the preset ratio range of 1:1 to 2:1. The random forest classification model is trained with the class-balanced training dataset.

[0012] Preferably, the content data of the ten metabolites were obtained by detection using an ultra-high performance liquid chromatography-tandem mass spectrometry system.

[0013] The beneficial effects of this invention are as follows: the combination of ten metabolites as biomarkers can synergistically characterize the adaptability of mung beans to dryland cultivation through relative content differences, and the reliability of multi-metabolite synergistic determination is better than that of single-metabolite determination; the standardization process uses the mean content of irrigated mung bean populations as a benchmark to eliminate systematic errors, making the detection data of different batches and under different instrument conditions comparable; the normalization weights are determined based on the variable importance of the random forest classification model, so that metabolites that contribute more to the determination of adaptability to dryland cultivation occupy a higher weight in the comprehensive score; the comprehensive score integrates the synergistic information of ten metabolites with a single quantitative index, and the determination process is objective and repeatable; the preset threshold is determined based on the 10th percentile of the comprehensive score distribution of known dryland-cultivated mung bean samples, and the determination boundary is clear; the random forest classification model is robust in classification performance through independent evaluation of training and test sets and stratified cross-validation; the class balancing process eliminates model bias caused by unbalanced sample numbers; ultra-high performance liquid chromatography-tandem mass spectrometry detection enables accurate and simultaneous determination of ten metabolites, providing a reliable data foundation for screening methods.

[0014] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0015] Figure 1 The mung bean morphology obtained by irrigation cultivation IC and dryland cultivation DC is an embodiment of the present invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the embodiments and accompanying drawings, so that those skilled in the art can implement it based on the description. Unless otherwise specified, the experimental methods used in the following embodiments are conventional methods; the reagents and materials used, unless otherwise specified, are commercially available.

[0017] In this invention, dryland cultivation refers to a cultivation method where mung beans are grown entirely without artificial irrigation, relying solely on natural rainfall. During the experimental period, the natural rainfall during the mung bean growth period was less than 60% of the local average rainfall for the same period over many years. Irrigated cultivation refers to a cultivation method where mung beans are artificially irrigated according to the local conventional irrigation system throughout their entire growth period, with no fewer than four irrigations throughout the entire growth period and a minimum irrigation volume of 40 m³ per irrigation. 3 / acre. Apart from the difference in water supply conditions, the two cultivation methods maintain the same agronomic conditions, including soil type, fertilization level, sowing density, and field management measures.

[0018] Example 1: Determination of biomarker combinations.

[0019] Field trials were conducted at a dryland farming experimental site in Xinjiang in early May 2022 and 2023, using 32 mung bean germplasm resources. Two cultivation environments were set up for each germplasm resource: irrigated cultivation and dryland cultivation, resulting in a total of 128 samples. Mung beans were harvested at maturity. The morphology of mung beans obtained from irrigated cultivation (IC) and dryland cultivation (DC) is as follows: Figure 1 As shown. Seeds from 5 plants were randomly selected from each germplasm sample under each cultivation environment and mixed. The seeds were threshed and cleaned, dried at 45°C, ground into powder using a cyclone mill, and passed through a 60-mesh sieve. The powder was then sealed and stored for later use.

[0020] Weigh appropriate amounts of each sample powder, add pre-cooled extraction solvent, vortex mix, and then extract ultrasonically at low temperature. After centrifugation, collect the supernatant, dry it under nitrogen, and reconstitute it to obtain the metabolomics analysis sample. Untargeted metabolomics analysis was performed using an ultra-high performance liquid chromatography-tandem mass spectrometry system. Chromatographic separation was performed using a reversed-phase column with gradient elution using an acetonitrile-water system containing formic acid as the mobile phase. Mass spectrometry used dual-mode scanning with positive and negative ions. Each sample underwent six extraction replicates. The quality control sample was a mixture of all sample extracts. Analysis was performed after every six injections during sequence analysis to ensure the stability of the detection data.

[0021] Peak identification, peak alignment, missing value imputation, and normalization were performed on the raw data obtained from the detection. A total of 75 metabolites were identified and classified into 8 chemical categories, including 37 flavonoids, 13 amino acids and their derivatives, 6 lignans and coumarins, 5 nucleotides and their derivatives, 3 alkaloids, 2 phenolic acids, 1 organic acid, and 8 other categories. Principal component analysis was performed on the metabolite data of 128 samples. The results showed that the irrigated cultivation samples and the dryland cultivation samples were significantly separated into two independent clusters. The first and second principal components explained 31.38% and 11.68% of the total variance, respectively, with a cumulative contribution of 43.06%, confirming that water supply conditions significantly altered the metabolite profile of mung beans. Further, an orthogonal partial least squares discriminant analysis model was established. The score plot showed clear separation between the two groups of samples, and the permutation test confirmed that the model had no overfitting risk. The model's cross-validation predictive goodness Q was [value missing]. 2 The value of 0.988 indicates that the model has good explanatory and predictive power.

[0022] Using the criteria of an absolute value greater than 2 for the base-2 logarithm of the difference, a p-value less than 0.05 for the significance test, and a variable projection importance (VIP) greater than 1.5, 22 significantly differentially expressed metabolites were identified. Of these, 20 were significantly upregulated in upland cultivated mung beans, including L-phenylalanine, L-proline, L-arginine, L-glutamic acid, L-tyrosine, tryptophan, 5-hydroxytryptophan, guanosine, apigenin, and various flavonoid glycosides. Only 4-hydroxyacetophenone and 4-glucosylvitexin were downregulated in upland cultivated mung beans. These differentially expressed metabolites are mainly involved in drought adaptation pathways such as osmotic regulation, antioxidant defense, and energy metabolism.

[0023] The content of 22 differentially expressed metabolites was used as the feature vector, and the cultivation environment category of the samples was used as the label. 70% of the samples were randomly selected as the training set and 30% as the test set. Random forest and support vector machine (SVM) classification models were constructed for these purposes. The random forest model achieved classification accuracies of 100% and 97.18% on the training and test sets, respectively, while the SVM model achieved 100% and 94.36% accuracy on the same sets. The generalization performance of the random forest model was superior to that of the SVM model. Based on the variable importance ranking in the random forest model, the top 10 contributing metabolites were selected as a biomarker combination for screening the adaptability of mung beans to dryland cultivation. This combination consists of L-tyrosine, L-phenylalanine, apigenin, L-glutamic acid, D-galactose, tryptophan, 5-hydroxytryptophan, guanosine, L-arginine, and L-proline, ranked from highest to lowest variable importance as follows: L-tyrosine, L-phenylalanine, apigenin, L-glutamic acid, D-galactose, tryptophan, 5-hydroxytryptophan, guanosine, L-arginine, and L-proline. The relative content differences of these 10 metabolites in dryland-cultivated and irrigated mung beans were used to synergistically characterize the dryland cultivation adaptability of mung beans.

[0024] Example 2: Construction of a random forest prediction model for quality breeding of mung beans grown in dryland.

[0025] 128 mung bean samples from known cultivation environments were collected. The contents of 10 metabolites, namely L-tyrosine, L-phenylalanine, apigenin, L-glutamic acid, D-galactose, tryptophan, 5-hydroxytryptophan, guanosine, L-arginine, and L-proline, in each sample were determined according to the method described in Example 1. The contents of the 10 metabolites were used to construct a feature vector. The cultivation environment category of the sample was used as the label, with irrigated cultivation category recorded as 0 and dryland cultivation category recorded as 1, to construct a training dataset.

[0026] After constructing the training dataset, the number of irrigated cultivation samples and the number of dryland cultivation samples in the training dataset are counted, and the ratio of the number of samples in the two classes is calculated. When the ratio is greater than 2:1 or less than 1:1, class balancing is performed on the training dataset. Class balancing uses synthetic minority class oversampling technology to generate synthetic samples for the class with fewer samples, so that the ratio of the two classes falls within the preset range of 1:1 to 2:1. In this embodiment, the number of irrigated cultivation samples and dryland cultivation samples are equal, and the ratio is 1:1, so class balancing is not required.

[0027] Before training the random forest classification model, a grid search was used to optimize the hyperparameters of the random forest classification model, including the number of decision trees, the maximum depth, and the minimum number of samples required for splitting. The search range for the number of decision trees was 100 to 500, the search range for the maximum depth was 5 to 20, and the search range for the minimum number of samples required for splitting was 1 to 10. The average accuracy of 5-fold cross-validation was used as the basis for hyperparameter selection. The optimized hyperparameters were 300 decision trees, 10 maximum depth, and 2 minimum number of samples required for splitting.

[0028] Stratified 5-fold cross-validation was performed on the training dataset. The samples in the training dataset were stratified according to cultivation environment category, ensuring that the class ratio of irrigated to dryland cultivation samples in each training and validation fold remained consistent with the overall class ratio of the training dataset. A random forest classification model was trained on each fold, and its classification performance was evaluated on the corresponding validation fold. The average evaluation result of each fold was used as the cross-validation performance of the random forest classification model. In this embodiment, the average cross-validation accuracy of the random forest classification model was 96.41%. The trained random forest classification model was independently evaluated on the test set. The test set classification accuracy was 97.18%, sensitivity was 95.00%, and specificity was 97.50%. Both the cross-validation performance and the test set evaluation results met the preset requirements of classification accuracy not less than 90%, sensitivity not less than 85%, and specificity not less than 85%, confirming that the random forest classification model passed performance validation. It was then used as a random forest prediction model for dryland mung bean quality breeding, and used to screen unknown mung bean samples for dryland mung bean quality breeding.

[0029] Example 3: A method for early screening of adaptability of mung beans for dryland cultivation.

[0030] The content data of 10 metabolites, namely L-tyrosine, L-phenylalanine, apigenin, L-glutamic acid, D-galactose, tryptophan, 5-hydroxytryptophan, guanosine, L-arginine and L-proline, in the mung bean sample to be tested were obtained. The detection method was ultra-high performance liquid chromatography-tandem mass spectrometry system, and the detection conditions were the same as in Example 1.

[0031] Using the mean content of corresponding metabolites in the irrigated mung bean population at the same experimental site as the benchmark, the content data of 10 metabolites were standardized. The standardization method was as follows: the ratio between the content of each metabolite in the mung bean sample and the mean content of the corresponding metabolite in the irrigated mung bean population was calculated, and the logarithmic transformation of this ratio to base 10 was used as the standardized relative content of each metabolite. When the content of a metabolite in the mung bean sample was higher than the mean of the irrigated population, the standardized relative content was positive; when the content was lower than the mean of the irrigated population, the standardized relative content was negative.

[0032] Normalized weights were calculated based on the variable importance of each metabolite in the random forest classification model constructed in Example 2. The normalized weights were calculated as follows: the sum of the variable importance of the 10 metabolites was used as the denominator, and the variable importance of each metabolite was used as the numerator to calculate the proportion of variable importance of each metabolite, thus obtaining the normalized weight. In this example, the normalized weights of the 10 metabolites were L-tyrosine 0.13, L-phenylalanine 0.12, apigenin 0.11, L-glutamic acid 0.10, D-galactose 0.10, tryptophan 0.09, 5-hydroxytryptophan 0.09, guanosine 0.09, L-arginine 0.09, and L-proline 0.08, with a sum of normalized weights of 1.00.

[0033] A weighted summation of the standardized relative contents was performed using the normalized weights of each metabolite to construct a comprehensive score for the mung bean sample to be tested. The comprehensive score was calculated as the sum of the products of the normalized weights of each metabolite and their corresponding standardized relative contents. The preset threshold was determined based on the 10th percentile of the comprehensive score distribution of known dryland-cultivated mung bean samples. In this embodiment, based on the comprehensive score distribution of 90 dryland-cultivated mung bean samples with known cultivation environments, the 10th percentile of the comprehensive score distribution, 0.10, was taken as the preset threshold. Samples with a comprehensive score lower than 0.10 were judged as irrigated cultivation adaptability samples, and samples with a comprehensive score not lower than 0.10 were judged as dryland cultivation adaptability samples.

[0034] Taking a mung bean sample from a dryland cultivation environment as an example, the standardized relative contents of 10 metabolites in the sample were measured as follows: L-tyrosine 0.42, L-phenylalanine 0.35, apigenin 0.28, L-glutamic acid 0.31, D-galactose 0.22, tryptophan 0.36, 5-hydroxytryptophan 0.25, guanosine 0.30, L-arginine 0.33, and L-proline 0.48. The comprehensive score of the sample was calculated by weighted summation according to the above normalized weights, which was 0.33. The comprehensive score of 0.33 is not lower than the preset threshold of 0.10, and the mung bean sample is judged to have breeding value for quality of dryland cultivated mung beans. Taking a mung bean sample from an irrigated cultivation environment as an example, the standardized relative contents of 10 metabolites in the sample were measured as follows: L-tyrosine 0.02, L-phenylalanine -0.01, apigenin -0.03, L-glutamic acid 0.01, D-galactose -0.05, tryptophan -0.02, 5-hydroxytryptophan -0.04, guanosine 0.00, L-arginine -0.03, and L-proline 0.04. The weighted summation yielded a comprehensive score of -0.01 for the sample. Since the comprehensive score of -0.01 is lower than the preset threshold of 0.10, the mung bean sample was determined to lack breeding value for quality of mung beans grown in dryland areas.

[0035] Comparative Example 1: Method for determining a single metabolite.

[0036] Samples were collected and metabolite contents were detected using the method described in Example 1. The standardized relative contents of individual metabolites, such as L-proline, L-tyrosine, L-phenylalanine, apigenin, and 5-hydroxytryptophan, were used as criteria. The suitability of a sample for dryland cultivation was directly determined by the relationship between the standardized relative contents of each individual metabolite and a preset threshold. The threshold was determined by the 10th percentile of the standardized relative contents of the individual metabolite in known dryland-cultivated mung bean samples. The remaining data processing methods were the same as in Example 3.

[0037] Comparative Example 2: Traditional physiological indicator screening methods.

[0038] Samples were collected according to the method described in Example 1. The protein content of each sample was determined using the Kjeldahl method, the total starch content was determined using the levorotatory method, and the total flavonoid content was determined using conventional methods. Using protein content, starch content, and total flavonoid content as screening indicators, a classification model based on these three physiological indicators was established using a logistic regression classifier. This model was then used to determine the adaptability of the samples to dryland cultivation and irrigated cultivation. The data partitioning and cross-validation methods for model training and evaluation were the same as in Example 2.

[0039] Comparative Example 3: Support Vector Machine Model Method.

[0040] Samples were collected and the contents of 10 metabolites were detected according to the method described in Example 1. The contents of the 10 metabolites were used as feature vectors. A support vector machine classification model was used instead of a random forest classification model to perform the discrimination between dryland cultivation adaptability and irrigated cultivation adaptability. The support vector machine used a radial basis kernel function. The penalty coefficient and kernel function coefficient were optimized by grid search. The rest of the data processing methods were the same as in Example 2.

[0041] Comparative Example 4: The method of combining all differential metabolites.

[0042] Samples were collected according to the method described in Example 1. The content of all 22 differential metabolites identified in Example 1 was used to construct a feature vector. A random forest classification model was used to distinguish between dryland cultivation adaptability and irrigated cultivation adaptability. The model construction method was the same as in Example 2, and the number of features was 22.

[0043] Comparative Example 1: Random non-differential metabolite combination method.

[0044] Samples were collected according to the method described in Example 1. After removing 22 differentially differentiated metabolites identified in Example 1 from 75 metabolites, 10 metabolites were randomly selected from the remaining 53 non-significantly different metabolites to form a random combination. The content of the 10 metabolites in the random combination was used to form a feature vector. A random forest classification model was used to distinguish between dryland cultivation adaptability and irrigated cultivation adaptability. The model construction method was the same as in Example 2.

[0045] Compare with Example 2: Unstandardized direct determination method.

[0046] Samples were collected and the contents of 10 metabolites were detected according to the method described in Example 1. The contents data were not standardized based on the mean contents of the irrigated mung bean population. The original contents data of the 10 metabolites were used directly to construct a feature vector to build a random forest classification model. The other model construction methods were the same as in Example 2, and the judgment methods were the same as in Example 3.

[0047] Experimental Example 1: Comparison of predictive performance of different combinations of metabolites.

[0048] Following the methods described in Examples 2 and 3, the classification performance comparison experiment was conducted on the same batch of 128 samples using the 10 metabolite biomarker combinations determined in Example 1, the single metabolite in Comparative Example 1, the traditional physiological indicators in Comparative Example 2, the support vector machine model in Comparative Example 3, the full 22 differential metabolite combinations in Comparative Example 4, the random 10 non-differential metabolite combinations in Control Example 1, and the unstandardized data in Control Example 2. 70% of the samples were randomly selected as the training set and 30% as the test set.

[0049] The evaluation metrics and calculation methods for each classification result are as follows: Accuracy equals the number of correctly classified samples divided by the total number of samples in the test set; Sensitivity equals the number of samples correctly classified as adaptable to dryland cultivation divided by the total number of dryland cultivation samples; Specificity equals the number of samples correctly classified as adaptable to irrigated cultivation divided by the total number of irrigated cultivation samples; F1 value is the harmonic mean of accuracy and sensitivity, where accuracy equals the number of samples classified as adaptable to dryland cultivation that are actually dryland cultivation samples divided by the total number of samples classified as adaptable to dryland cultivation; The area under the receiver operating characteristic curve (AUC) is calculated based on the ranking of the discrimination scores of each method's output on the test set samples.

[0050] The test set contained 39 samples, including 20 from dryland cultivation and 19 from irrigated cultivation. In the classification results of the random forest model combined with the 10 metabolite biomarker combinations identified in Example 1, the judgment results for 38 samples were consistent with the actual cultivation environment, with an accuracy of 97.44%. Among these, 19 samples from dryland cultivation were correctly classified (sensitivity 95.00%), and 19 samples from irrigated cultivation were correctly classified (specificity 100%). The support vector machine model correctly classified 36 samples with an accuracy of 92.31%. The method using all 22 differentially expressed metabolites correctly classified 35 samples with an accuracy of 89.74%. The combination of the top 5 most important metabolites correctly classified 33 samples with an accuracy of 84.62%. The single L-proline determination method correctly classified 26 samples with an accuracy of 66.67%. The traditional physiological indicator screening method correctly classified 25 samples with an accuracy of 64.10%. The random, non-differential metabolite combination method correctly classified 22 samples with an accuracy of 56.41%. The unstandardized direct judgment method correctly classified 31 samples with an accuracy of 79.49%.

[0051] Analysis of the above results shows that the accuracy, sensitivity, and specificity of the combination of 10 metabolite biomarkers are significantly higher than those of the comparative and control examples. The accuracy of the 10-metabolite combination is approximately 7.7 percentage points higher than the combination using all 22 differential metabolites, indicating that this invention does not necessarily improve performance with more features. Instead, it achieves a balance between feature simplification and performance improvement by filtering metabolites with little contribution to classification or even introducing noise through random forest variable importance screening. This simplification process is difficult for those skilled in the art to anticipate. The accuracy of the 10-metabolite combination is approximately 5.1 percentage points higher than the support vector machine model, indicating that the random forest model combined with variable importance weighting can more fully utilize the collaborative discriminative information of the 10 metabolites. The accuracy of the 10-metabolite combination is approximately 30 percentage points higher than that of a single metabolite, confirming that the reliability of multi-metabolite collaborative discrimination is significantly better than that of single-metabolite discrimination. The accuracy of the 10 metabolite combinations was approximately 33 percentage points higher than that of traditional physiological indicator screening methods, confirming that the discriminative information at the metabolite level is far superior to conventional nutritional quality indicators such as protein, starch, and total flavonoids. This is because the total flavonoid content is mainly regulated by genetic factors, with water supply having little effect. Protein and starch contents exhibit large variations among germplasms and are highly correlated with the genetic background of the variety, making it difficult to reliably distinguish between cultivation environments. The accuracy of random, non-differential metabolite combinations was only 56.41%, close to random guessing, indicating that the determination of the 10 metabolite combinations in this invention was not arbitrary but must be based on differential metabolite analysis and random forest variable importance screening. The accuracy of the unstandardized direct determination method decreased by approximately 18 percentage points compared to the standardized method, confirming that standardization based on the mean content of irrigated mung bean populations can eliminate systematic errors between samples from different batches and under different testing conditions, and is a necessary technical means to ensure the accuracy of the determination.

[0052] Experiment Example 2: Cross-Year Verification Experiment.

[0053] To verify the universality of the method of the present invention across different years, samples collected in 2022 were used as the training set, and samples collected in 2023 were used as the test set. A random forest classification model was constructed and performed according to the methods described in Examples 2 and 3. The test set consisted of 64 samples, including 32 samples from dryland cultivation and 32 samples from irrigated cultivation. The judgment results for 60 samples were consistent with the actual cultivation environment, with an accuracy of 93.75%, a sensitivity of 90.63%, and a specificity of 96.88%. The experiment was repeated using samples collected in 2023 as the training set and samples collected in 2022 as the test set, achieving an accuracy of 92.19%, a sensitivity of 87.50%, and a specificity of 96.88%. Both cross-year verification schemes showed an accuracy of no less than 90%, a sensitivity of no less than 85%, and a specificity of no less than 85%, demonstrating that the method of the present invention and the method for determining the preset threshold have good stability and universality across different years, can withstand climate fluctuations between years, and are suitable for the practical needs of continuous screening over multiple years in breeding practice.

[0054] Experiment Example 3: Cross-variety and cross-regional verification experiment.

[0055] To verify the applicability of the method of the present invention to varieties not involved in modeling, 6 germplasm resources were randomly selected from 32 mung bean germplasm resources. All 12 samples from these 6 resources under irrigated and dryland cultivation conditions were removed from the training dataset and used only for model validation. The remaining 116 samples were used to construct a model and determine thresholds according to the methods described in Examples 2 and 3. Subsequently, the 12 selected samples were evaluated. The evaluation results for 11 samples were consistent with the actual cultivation environment, with an accuracy rate of 91.67%. To further verify the regional universality of the method of the present invention, 10 of the aforementioned 32 germplasm resources were planted at another dryland cultivation experimental site in Xinjiang Uygur Autonomous Region using the same dryland and irrigated cultivation operations. Samples were collected at maturity according to the method described in Example 1, and the content of 10 metabolites was detected. The model constructed in Example 2 was used to evaluate the samples. The evaluation results for 18 out of 20 samples were consistent with the actual cultivation environment, with an accuracy rate of 90.00%. The above results show that the method of the present invention can accurately determine the mung bean varieties that did not participate in model training and samples from different experimental sites, indicating that the combination of 10 metabolite biomarkers and their thresholds has universality across varieties and regions, and can meet the application needs of large-scale, multi-source germplasm material early screening in breeding practice.

[0056] Experiment Example 4: Correlation verification experiment between comprehensive score and quality index.

[0057] To verify the ability of the comprehensive score of this invention to characterize the quality breeding value of mung beans, all 64 samples from 2023 were used as subjects. The comprehensive score of each sample was calculated according to the method described in Example 3. Simultaneously, the protein content of each sample was determined using the Kjeldahl method. The Pearson correlation coefficient between the comprehensive score and protein content was calculated, and the result was 0.71, indicating a highly significant correlation. Statistical analysis was conducted on dryland cultivated samples with a comprehensive score not lower than the threshold of 0.10. Their average protein content was 29.31 g / 100 g, significantly higher than the 26.98 g / 100 g of irrigated cultivated samples with a comprehensive score lower than the threshold. These results indicate that the comprehensive score constructed in this invention can not only accurately determine the adaptability of mung beans to dryland cultivation, but also has a positive correlation with the protein accumulation level of mung beans under dryland cultivation conditions. Samples with higher comprehensive scores correspond to higher protein content, directly reflecting the quality breeding value of dryland cultivated mung beans. This demonstrates that the screening results of this invention are supported by real quality connotations, rather than simply statistical classification results.

[0058] Experiment Example 5: Experiment to verify the rationality of threshold setting.

[0059] To verify the rationality of using the 10th percentile of the comprehensive score distribution of dryland-cultivated mung bean samples as the preset threshold, based on the model and comprehensive score constructed in Example 3, the 5th percentile (0.05), 10th percentile (0.10), and 15th percentile (0.15) were used as alternative thresholds to re-perform the judgment on 39 samples in the test set and compare the classification performance. When using the 5th percentile threshold, the sensitivity for dryland-cultivated samples was 100%, but 4 samples from irrigated cultivation were misclassified as dryland-cultivation adaptability, reducing the specificity to 78.95%. When using the 15th percentile threshold, the specificity was 100%, but 5 samples from dryland-cultivated samples were missed, reducing the sensitivity to 75.00%. When using the 10th percentile threshold, the sensitivity was 95.00%, the specificity was 100%, and the accuracy was 97.44%, making it the optimal solution among the three alternative thresholds. The above results indicate that setting the preset threshold as the 10th percentile of the comprehensive score distribution of known dryland mung bean samples can effectively suppress misjudgment of irrigated samples while ensuring a high detection rate of dryland adaptability samples, achieving the best balance between sensitivity and specificity. This threshold determination method has clear operational basis and stable technical effect.

[0060] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the present invention's vehicle wading warning system will be readily apparent to those skilled in the art.

[0061] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A combination of biomarkers for screening the adaptability of mung beans to dryland cultivation, characterized in that, It consists of the following ten metabolites: L-tyrosine, L-phenylalanine, apigenin, L-glutamic acid, D-galactose, tryptophan, 5-hydroxytryptophan, guanosine, L-arginine, and L-proline; the relative content differences of these ten metabolites in dryland-cultivated and irrigated mung beans were used to synergistically characterize the dryland cultivation adaptability of mung beans.

2. A method for early screening of the adaptability of mung beans for dryland cultivation, characterized in that, Includes the following steps: The content data of ten metabolites in the mung bean samples to be tested were obtained. The ten metabolites are L-tyrosine, L-phenylalanine, apigenin, L-glutamic acid, D-galactose, tryptophan, 5-hydroxytryptophan, guanosine, L-arginine, and L-proline. The content data of the ten metabolites were standardized based on the mean content of the corresponding metabolites in the irrigated mung bean population to obtain the standardized relative content of each metabolite. The normalized weights were calculated according to the importance of each metabolite in the random forest classification model. The standardized relative contents were then weighted and summed using the normalized weights to construct the comprehensive score of the mung bean samples to be tested. When the comprehensive score is not lower than a preset threshold, the mung bean samples to be tested are determined to have breeding value for dryland cultivation and are suitable for dryland cultivation. The preset threshold is determined based on the 10th percentile of the comprehensive score distribution of known dryland cultivated mung bean samples. Samples below the 10th percentile are considered to be suitable for irrigated cultivation, and samples above the 10th percentile are considered to be suitable for dryland cultivation.

3. The method for early screening of mung bean adaptability for dryland cultivation as described in claim 2, characterized in that, The method for constructing a random forest classification model includes the following steps: collecting mung bean samples from multiple known cultivation environments, measuring the content of ten metabolites in each sample (L-tyrosine, L-phenylalanine, apigenin, L-glutamic acid, D-galactose, tryptophan, 5-hydroxytryptophan, guanosine, L-arginine, and L-proline), and using the content of these ten metabolites to construct a feature vector; constructing a training dataset using the cultivation environment category of the samples as a label; training a random forest classification model using the training dataset; evaluating the classification performance of the random forest classification model, and using a random forest classification model whose classification performance meets the preset requirements as a random forest prediction model for quality breeding of mung beans grown in dryland areas, and performing quality breeding screening for unknown mung bean samples grown in dryland areas.

4. The method for early screening of mung bean adaptability for dryland cultivation as described in claim 3, characterized in that, The standardization process uses the mean content of corresponding metabolites in the irrigated mung bean population as the benchmark. The ratio between the content of each metabolite in the mung bean sample to be tested and the corresponding mean is calculated. The logarithmic transformation result of the ratio to base 10 is then used as the standardized relative content.

5. The method for early screening of mung bean adaptability for dryland cultivation as described in claim 3, characterized in that, Before training the random forest classification model, a grid search is used to optimize the hyperparameters of the random forest classification model, including the number of decision trees, maximum depth, and minimum number of samples required for splitting. The average accuracy of cross-validation is used as the basis for hyperparameter selection, and the optimized hyperparameters are used to train the random forest classification model.

6. The method for early screening of mung bean adaptability for dryland cultivation as described in claim 5, characterized in that, Stratified cross-validation was performed on the training dataset. The samples in the training dataset were stratified according to the cultivation environment category, so that the class ratio of irrigated cultivation samples to dryland cultivation samples in the training and validation folds of each cross-validation fold was consistent with the overall class ratio of the training dataset. The random forest classification model was trained on each fold and the classification performance was evaluated on the corresponding validation fold. The average value of the evaluation results of each fold was used as the cross-validation performance of the random forest classification model. The classification performance of the random forest classification model was double-validated based on the cross-validation performance and the evaluation results of the test set. When both met the requirements of classification accuracy of not less than 90%, sensitivity of not less than 85%, and specificity of not less than 85%, the random forest classification model was confirmed to have passed the performance validation.

7. The method for early screening of mung bean adaptability for dryland cultivation as described in claim 3, characterized in that, After constructing the training dataset, the number of irrigated cultivation samples and the number of dryland cultivation samples in the training dataset are counted, and the ratio of the number of samples of the two classes is calculated. When the ratio is greater than 2:1 or less than 1:1, class balancing is performed on the training dataset. The class balancing process uses synthetic minority class oversampling technology to generate synthetic samples for the class with fewer samples, so that the ratio of the number of samples of the two classes falls within the preset ratio range of 1:1 to 2:

1. The random forest classification model is trained using the class-balanced training dataset.

8. The method for early screening of mung bean adaptability for dryland cultivation as described in claim 2, characterized in that, The content data of ten metabolites were obtained by detection using an ultra-high performance liquid chromatography-tandem mass spectrometry system.

Images