JP2023090795A - Method of predicting soybean yield - Google Patents

Abstract

To provide a method of predicting soybean yield at an early stage with high accuracy.SOLUTION: In a method of predicting soybean yield, analysis data on one or more components is obtained from a leaf sample collected from soybean, and the soybean yield is predicted by using a correlation between the data and the soybean yield.SELECTED DRAWING: None

Description

The present invention relates to a method for early prediction of soybean yield.

Soybean is an important cereal and is widely eaten in Japan and around the world. In addition, unlike other representative grains such as rice, wheat and corn, it has a high protein and lipid ratio and is rich in nutritional value. Therefore, it is also important as a feed and oil material, and techniques for increasing the yield are being developed.
The growing period of soybean varies slightly depending on the variety and cultivation conditions, but it usually takes a long period of 4 to 5 months from sowing to harvest. Therefore, in the development of techniques for increasing the yield of soybeans, much time is required for cultivation in order to evaluate the yield. Furthermore, in the seasonal and climatic conditions of Japan, soybeans, which require 4-5 months to harvest, are generally cultivated once a year. Yield evaluation in outdoor cultivation can only be done once a year, which is an obstacle to the development of yield-increasing technology.
A method for predicting yield at an early stage has been sought. Also, in the actual production scene, if the yield can be predicted at an early stage, the producer can easily decide whether to introduce costly additional technology to secure a stable yield.

Various methods have been studied so far for evaluating the yield property at an early stage using the correlation between the growth state of the growing plant body and the yield. For example, in Non-Patent Document 1, a method using the correlation between the main stem length and the yield (r = 0.51) measured about 40 days after sowing soybeans, and in Non-Patent Document 2,
A method is disclosed that utilizes the correlation (r=0.66) between the dry weight of above-ground parts measured about 60-70 days after sowing and the yield. In addition, in Non-Patent Documents 3 and 4, using diagnostic imaging technology, NDVI (Normalized Vegetation Index), LAI (Leaf Area Index) and canopy spectral reflectance are measured in the field, and an attempt to evaluate growth and yield. is disclosed.

However, the method of Non-Patent Document 1 has the possibility of predicting the yield relatively early, but the correlation is not sufficient, and the method of Non-Patent Document 2 improves the correlation, but the prediction time is the start of cultivation. It takes more than 2 months and half of the growth period has passed since the beginning of the experiment, and it is invasive to measure the above-ground dry weight. . The methods of Non-Patent Documents 3 and 4 can be said to be non-destructive and simple measurements.

In addition, in rice, metabolites extracted from the aerial part about 15 days after seeding were analyzed by GC-MS.
It has been reported that a hybrid rice yield prediction model was created using these data by comprehensively measuring the data (Non-Patent Document 5). Predictive evaluation of the model called validation has not been performed, and verification cannot be said to be sufficient. In addition, it is invasive and not suitable for evaluation when it is desired to match predictive factors and yields for each individual.

Yoichi Fujita et al., 2009 ``Kanto Tokai Hokuriku Agriculture'' research results information, ``Yield components and growth indicators based on high-yield cases of soybean ``Enrei'' in heavy viscosity areas, http://www.naro.affrc .go.jp/org/narc/seika/kanto21/12/21_12_04.html Kenichi Inoue, Masaki Takahashi, Abstracts of the 229th Annual Meeting of the Crop Science Society of Japan, 2010, p50, "High Yield Growth Phase of Soybean Seen from Substance Production and Nitrogen Accumulation" Tomoya Osanami et al., Abstracts of the 245th Annual Meeting of the Crop Science Society of Japan, 2018, p83, "Simple pre-flowering growth diagnosis technology for soybean" Tomoya Watanabe et al., Abstracts of the 245th Annual Meeting of the Crop Science Society of Japan, 2018, p84, "Yield evaluation of soybean using non-destructive measurement and convolutional neural network" Dan, Z. et al., Scientific Reports, 2016, 6, 21732

The present invention relates to providing a method for early and accurate prediction of soybean yield.

As a result of various studies on the yield evaluation of soybeans, the present inventors found that metabolites contained in leaves have components whose abundance correlates with yield, and that as early as about one month after sowing It was found that the final yield can be evaluated at the individual level by collecting one expanded leaf and analyzing the components contained in the leaf.

That is, the present invention provides a soybean yield prediction method, which obtains analytical data of one or more components from a leaf sample collected from soybean, and predicts the soybean yield using the correlation between the data and the soybean yield. ,I will provide a.

According to the method of the present invention, soybean yield can be predicted early. As a result, for example, it becomes easier to decide whether to introduce additional technologies to secure yields, and it is possible to significantly improve the efficiency of the development of yield-increasing technologies.

FIG. 10 is a diagram showing the relationship between yield prediction values and actual yield values based on an OPLS model constructed using all 125 data. FIG. 10 is a diagram showing the relationship between the yield prediction value and the actual measurement value by a machine learning model constructed using all 125 data. The figure which shows the relationship between each predicted value and measured value of data for learning and data for verification. All 125 data matrices were randomly divided into two groups (for learning and for verification), one of the 63 matrix data was used for learning and the remaining 62 matrix data were used for verification. Analysis data of all components below VIP value 11th in the model in Figure 1, analysis data of all components below 21st, analysis data of all components below 31st, and analysis data of all components below 351st FIG. 10 is a diagram showing the R ² (indicated as R2Y in the figure) and Q ² (indicated as Q2 in the figure) values of each model constructed by the OPLS method. Of the analysis data of the top 1 to 10 VIP value components in the model in Figure 1, ^each model constructed by the OPLS method for any two combinations (45 ways) R ) and Q ² (indicated as Q2 in the figure) values. Each constructed by the OPLS method using the analysis data of the top 1st and 2nd, 11th and 12th, 21st and 22nd, . A diagram showing the R ² (labeled as R2Y in the figure) and Q ² (labeled as Q2 in the figure) values of the model. The components of the top 1st, 2nd and 3rd, 11th, 12th and 13th, 21st, 22nd and 23rd, . FIG. 4 is a diagram showing the R ² (indicated as R2Y in the figure) and Q ² (indicated as Q2 in the figure) values of each model constructed by the OPLS method using analytical data. VIP value top 1st, 2nd, 3rd and 4th, 11th, 12th, 13th and 14th, 21st, 22nd, 23rd and 24th, . and the R ² (indicated as R2Y in the figure) and Q ² (indicated as Q2 in the figure) values of each model constructed by the OPLS method using the analytical data of the components at positions 222, 223 and 224. illustration. Each constructed by the OPLS method using the analysis data of the top 1st to 5th, 11th to 15th, 21st to 25th, . A diagram showing the R ² (labeled as R2Y in the figure) and Q ² (labeled as Q2 in the figure) values of the model. Each constructed by the OPLS method using the analysis data of the top 1st to 6th, 11th to 16th, 21st to 26th, . A diagram showing the R ² (labeled as R2Y in the figure) and Q ² (labeled as Q2 in the figure) values of the model. Each constructed by the OPLS method using the analysis data of the top 1st to 7th, 11th to 17th, 21st to 27th, . A diagram showing the R ² (labeled as R2Y in the figure) and Q ² (labeled as Q2 in the figure) values of the model. Each constructed by the OPLS method using the analysis data of the top 1st to 8th, 11th to 18th, 21st to 28th, . A diagram showing the R ² (labeled as R2Y in the figure) and Q ² (labeled as Q2 in the figure) values of the model. Each constructed by the OPLS method using the analysis data of the top 1st to 9th, 11th to 19th, 21st to 29th, . A diagram showing the R ² (labeled as R2Y in the figure) and Q ² (labeled as Q2 in the figure) values of the model. Each constructed by the OPLS method using the analysis data of the top 1st to 10th, 11th to 20th, 21st to 30th, . A diagram showing the R ² (labeled as R2Y in the figure) and Q ² (labeled as Q2 in the figure) values of the model. The figure which shows the relationship between the predicted value of a yield by the OPLS model (model A) constructed using the analysis data of 100 components per 1 data, and an actual value. FIG. 10 is a diagram showing the results of yield prediction using model A for test plots 1 to 10 (difference from test plot 1). The figure which shows the field prediction yield using the model A in the MIX manure application plot. The figure which shows the field yield in MIX compost application. Figure 2 shows a comparison of predicted yield using Model A and observed yield at 2 weeks and 8 weeks after sowing. The figure which shows the relationship between the predicted value of a yield by the OPLS model constructed|assembled using field data (analysis data of 431 components per 1 data), and an actual value.

In the present invention, soybean means soybean (scientific name: Glycine max) which is an annual plant of the legume family. The varieties include Fukuyutaka, Enrei, Sato no Hohoemi, Yuagari Musume, Ryuho, Suzuyutaka, etc., but the present invention is not limited to them.

The growth stage from germination to defoliation of soybean is VC: primary leaf development stage (around 7 days after sowing)
, R1-2: flowering period (around 50 days after sowing), R3-4: seeding period (around 70 days after sowing), R
5-6: Divided into seed enlargement stage (around 90 days after sowing) (Fehr, WR, Caviness, CE,
1977. Stages of soybean development.
and Home Economics Experiment Station, Iowa State University, Ames, Iowa). In the present invention, the soybean leaves used as samples may be collected from the primary leaf development stage (VC) when the leaves can be collected to the grain hypertrophy stage (R5-6), preferably Primary leaf development stage to R3-4 stage, more preferably 14 days after sowing to R3-4 stage, more preferably 21 days after sowing to R1-2 stage, still more preferably 28 days after sowing to R1-2 stage is mentioned. The number of days before and after each growth stage is preferably within 10 days.
Alternatively, soybean leaves are harvested 7 days or more after sowing, preferably 14 days or more, more preferably 21 days or more, still more preferably 28 days or more, and preferably before 50 days after sowing. Preferably before 40 days after seeding, more preferably before 35 days. It may also be 7 to 50 days after seeding, preferably 14 to 40 days, more preferably 28 to 35 days. For example, it is suitable to harvest leaves from soybeans 30 days ± 3-5 days after sowing.

The part from which the leaf is collected is not particularly limited, but for example, the central compound leaf of the three compound leaves that constitute the true leaf that is one or two leaf years older than the uppermost true leaf can be collected.

In the present invention, the analytical data of the components to be obtained include high performance liquid chromatography (HPLC), gas chromatography (GC), ion chromatography, mass spectrometry (MS), near infrared spectroscopy (NIR), Fourier transform Infrared spectroscopy (FT-IR), nuclear magnetic resonance spectroscopy (NMR), Fourier transform nuclear magnetic resonance spectroscopy (FT-NMR), inductively coupled plasma mass spectrometer (ICP-MS), liquid chromatograph and mass spectrometry Combined LC/MS
Data analyzed and measured using instrumental analysis means such as are preferably mass spectrometric data, more preferably mass spectrometric data by LC/MS.
The mass spectrometry data includes accurate mass (“m/z value”), ion intensity, retention time, etc., but is preferably accurate mass information.

In order to apply the leaf sample to the instrumental analysis method, it is appropriately pretreated according to the analysis method. Usually, the collected leaves are wrapped in aluminum foil and immediately frozen in liquid nitrogen to stop the metabolic reaction. , dried by freeze-drying, and then subjected to extraction operations.
Extraction is performed by pulverizing a freeze-dried leaf sample using a bead pulverizer or the like, then adding an extraction solvent and stirring. Extraction solvents used here include methanol, ethanol, butanol, acetonitrile, chloroform, ethyl acetate, hexane, acetone, isopropanol, water, and mixtures thereof. When LC/MS is used as an analysis means, an 80 v/v % aqueous solution of methanol containing an internal standard is preferably used.

In the present invention, the components in the leaves to be analyzed include soybean metabolites separated and detected by LC/MS. Preferably, the exact mass provided by mass spectrometry (m/
Components where z) is from 139 to 1156 are included. More preferably, the 431 components described in Tables 1a to 1c below, defined by accurate mass (m/z value) provided by mass spectrometry, are included. In the process of separation and detection by LC/MS, when different molecular ion peaks of partial degradation products and adducts (M + H, M + Na, etc.) are generated from metabolites, the detected partial degradation products are different from the original metabolites. be a component of

The 431 components were selectively extracted from soybean metabolites, and the selection method is shown in detail in the examples. Cultivate 125 soybean strains with different strains, 2) collect one leaf about 1 month after sowing, 3) extract components using 80 v / v% methanol aqueous solution, 4)
LC/MS analysis was performed to obtain molecular ion information (accurate mass, m/z) and structural information derived from fragments, 5) component derived peaks were extracted, and then alignments were performed to align each peak across each sample. Processing, isotope peak removal, inter-sample peak intensity correction, and noise removal are performed to obtain analytical data for 431 components. The method for peak intensity correction between samples is not particularly limited, but examples thereof include the pooled QC method and correction using an internal standard substance. In the pooled QC method, a sample called pooled QC is prepared by mixing a constant amount from all samples in the same batch, and pooled QC is analyzed at a constant frequency (about once every 5 to 9 times) between each sample. By performing the process of calculating the estimated value of ``what will happen to each peak intensity assuming that the QC sample was analyzed when each sample was analyzed'' and correcting it with that value This is to correct the sensitivity between each sample. Correction using an internal standard substance was performed by correcting the peak area value of an internal standard substance (lidocaine, 10-camphorsulfonic acid, etc.) added in the same amount to each sample, thereby correcting the sensitivity between samples. It is something to do.
The data correction method does not significantly affect the correlation with yield and the performance of the prediction model.

In addition, the results of correlation analysis between the analytical data of 431 components in the obtained 125 leaves and the corresponding yield data (the single correlation coefficient r between the peak area of the analytical data of each component and the yield and the test of uncorrelation p-values were calculated by ), and certain components were shown to be significantly correlated with yield (see Tables 3a-3f below).

Therefore, among the 431 components, as the component to be analyzed in the present invention, the component that has a significant correlation with the yield (p < 0.05) and the absolute value of the correlation coefficient |r|> 0.51, that is, component No. . 13, 14, 17, 20, 21, 22, 23, 28, 35, 36, 37, 39, 4
1, 42, 44, 47, 48, 51, 52, 54, 57, 58, 68, 71, 73, 80
, 85, 86, 90, 91, 96, 98, 99, 100, 107, 108, 110, 12
2, 125, 131, 134, 135, 137, 139, 142, 149, 150, 15
3, 157, 159, 160, 161, 171, 174, 176, 179, 181, 18
2, 188, 202, 208, 209, 214, 215, 217, 218, 228, 23
0, 235, 244, 245, 246, 247, 249, 251, 252, 253, 26
1, 264, 268, 275, 278, 279, 280, 282, 283, 284, 28
8, 294, 296, 298, 299, 305, 308, 310, 313, 317, 32
5, 327, 329, 330, 341, 347, 353, 355, 356, 363, 36
7, 369, 370, 384, 389, 395, 421, 422, 423, 428 and 4
It preferably contains one or more selected from 31. All of the above components had a VIP value of 1.16 or more, which will be described later. If the VIP value was 1.30 or more, the absolute value of the correlation coefficient |r|>0.51.

Furthermore, among the 431 components, as the component to be analyzed in the present invention, the correlation with the yield is significant (
p<0.05) and the absolute value of the correlation coefficient |r|>0.63, that is, component No. 1
4, 22, 23, 36, 37, 41, 42, 51, 52, 68, 90, 122, 139,
149, 159, 214, 228, 230, 235, 247, 249, 252, 253,
268, 275, 278, 284, 288, 298, 305, 308, 313, 317,
More preferably, it contains one or more selected from 329, 347, 363, 395, 421, 422 and 428. All the above components have a VIP value of 1.522 or more, which will be described later.
If it is 1.62 or more, the absolute value of the correlation coefficient |r|>0.63.

Furthermore, among the 431 components, as the component to be analyzed in the present invention, the correlation with the yield is significant (
p<0.05) and the absolute value of the correlation coefficient |r|>0.66, that is, component No. 1
4, 23, 36, 37, 41, 51, 68, 90, 122, 149, 214, 230, 2
35, 247, 249, 252, 275, 284, 298, 305, 308, 313, 3
More preferably, one or more selected from 17, 347, 363, 421, 422 and 428 are included. All of the above components had a VIP value of 1.59 or more, which will be described later, and if the VIP value was 1.652 or more, the absolute value of the correlation coefficient |r|>0.66.

In Tables 1a to 1c, 431 components are defined by accurate mass obtained by mass spectrometry, and the compositional formula of the compound can be estimated from these accurate mass data. Moreover, partial structural information of the compound can be obtained from the MS/MS data acquired at the same time as the analysis. Therefore, from the compositional formula and the partial structure information, the target component can be estimated, and those that can be compared with the reagent can be identified.

For example, among the 431 components, those identified as a result of analysis include the following. Component No. 10 is 4-coumaric acid, Component Nos. 68, 90, 122 and 308 are the same component, the composition formula of which is C ₂₁ H ₂₂ O ₁₁ , and the composition formula of the aglycon is C ₁₅ H ₁₂ O. Monoglucoside in which glucose is bound to the dehydroflavonol of _{No. 6} , Component No. 92 is prunin (naringenin 7-O-glucoside), Component No. 119 has the compositional formula C ₁₅ H ₁₀ O ₆
The flavonoid, Component No. 139, has a composition formula of C ₂₁ H ₂₂ O ₁₂ , and a monoglucoside in which glucose is bound to dehydroflavonol whose aglycone composition formula is C ₁₅ H ₁₂ O ₇ , Component No. 277 is a prenylated flavonoid of the composition formula C ₁₅ H ₁₀ O ₅ whose composition formula is C ₂₆ H ₃₀ O ₁₀ , component No. 295 is a composition formula C ₁₅
Diglycoside in which glucose and rhamnose are bound to the flavonol of H ₁₀ O ₆ , Component N
No. 296 and component No. 395 are the same component, the composition formula of which is shown by C ₃₃ H ₄₀ O ₁₉ , and the composition formula of the aglycon is C ₁₅ H ₁₀ O ₆ flavonol with one glucose and two rhamnose. The bound triglycoside, component No. 302 is a diglycoside, component No. 302, represented by the compositional formula C ₂₁ H ₃₆ O ₁₀ , in which glucose and arabinose are bound to geraniol.
429 was presumed to be soyasaponin βg. Also, from the coincidence with the reagent, component No. 76 is daidzein, component No. 89 is genistein, component No. 276 is genistin, and component No. 399.
was identified as malonylgenistin, and component No. 421 and component No. 422 are the same component and are soyasaponin Bb.

Among these, as the component to be analyzed in the present invention, preferably the correlation with the yield is significant (p <
0.05) and the absolute value of the correlation coefficient |r|>0.51, such as soyasaponin Bb;
A monoglucoside in which glucose is bound to dehydroflavonol having a composition formula of C ₂₁ H ₂₂ O ₁₁ and an aglycone composition formula of C ₁₅ H ₁₂ O ₆ ; a composition formula of C ₃₃ H ₄₀ O ₁₉
A triglycoside _{in which} one glucose and two rhamnose are bound to a flavonol of C ₁₅ H ₁₀ O _{6 ;} Examples include monoglucosides in which glucose is bound to dehydroflavonol, which is C ₁₅ H ₁₂ O ₇ .

The soybean yield prediction means is that the above 431 components, preferably the correlation with yield is significant (p <
0.05) and the absolute value of the correlation coefficient |r|>0.51, more preferably significant (p<
0.05) and the absolute value of the correlation coefficient |r|>0.63, more preferably significant (p<
0.05) and the absolute value of the correlation coefficient |r|>0.66, e.g. is also measured and the yield value can be estimated from the correlation between known yield and peak area.

In addition, the yield can be predicted by using a plurality of analytical data of the above 431 components and comparing them with a yield prediction model constructed using a multivariate analysis method.
That is, by collecting soybean leaves after a predetermined period of time from sowing, obtaining an analysis sample, subjecting the analysis sample to instrumental analysis to obtain instrumental analysis data, and comparing the instrumental analysis data with a yield prediction model , the yield of the soybean can be predicted.

The yield prediction model can be constructed by performing regression analysis using the peak area values of the corrected analytical data of the components having each accurate mass as the explanatory variables and the yield values as the objective variables. As a regression analysis method, for example, principal component regression analysis, PLS (Partial least squares projection t
o latent structures) regression analysis, OPLS (Orthogonal projections to latent struct
ures) regression analysis, generalized linear regression analysis, bagging, support vector machine, random forest, neural network regression analysis and other machine learning/regression analysis techniques, and other multivariate regression analysis techniques. Among these, the PLS method, the OPLS method which is an improved version of the PLS method,
Alternatively, it is preferable to use machine learning/regression analysis techniques. The OPLS method has the same predictability as the PLS method, but is superior for purposes such as the present in that it is easier to visualize for interpretation. Both the PLS method and the OPLS method are methods of aggregating information from high-dimensional data, replacing it with a small number of latent variables, and using the latent variables to express objective variables. It is important to select an appropriate number of latent variables, and cross-validation is often used to determine the number of latent variables. That is, divide the model building data into several groups, use some groups for model validation and other groups for model building to estimate the forecast error, and repeat this process while switching groups to obtain the forecast error. The number of latent variables with the smallest sum is chosen.

Evaluation of the prediction model is mainly judged by two indexes. One is R ² representing prediction accuracy and the other is Q ² representing predictability. ^R2 is the square of the correlation coefficient between the measured values of the data used to construct the prediction model and the predicted values calculated by the model, and the closer to 1, the higher the prediction accuracy. On the other hand, ^Q2 is the result of the above cross-validation, and represents the square of the correlation coefficient between the measured value and the predicted value, which is the result of repeated model verification. In the soybean yield prediction model of the present invention, it is preferable to use Q ² >0.50 as a criterion for model evaluation. Since R ² >Q ² is always satisfied, Q ² >0.50 simultaneously satisfies R ² >0.50.
Below, various soybean yield prediction models were created using the peak area values of the analysis data of all or part of the above 431 components and the grain yield, and the results of verifying their accuracy are shown. Of these, the OPLS model can preferably use the model that satisfies Q ² >0.50.

(1) Construction of yield prediction model using all 431 component information OPLS model from all 125 data matrices with peak area values and yield values of analytical data of 431 components per data (Fig. 1) built. In addition, when constructing, the peak area value and yield data of the analysis data of each component are autoscaled to an average of 0 and a variance of 1
converted to With R ² =0.87 and Q ² =0.78, the model can be said to have high prediction performance.
VIP (Variable Importance in the P
The contribution to model performance given to each component, called projection (variable importance in projection) value, is calculated.
The VIP value is obtained by Equation 1 below.

The larger the VIP value, the greater the degree of contribution to the model, and the absolute value of the correlation coefficient is also correlated. A list of VIP values is shown in Tables 4a-4f below.

(2) Visual Mining Studio (hereinafter referred to as VMS,
NTT Data Mathematical Systems Co., Ltd.) was used.
(2-1) A model (random forest) that has the peak area value and yield value of the analysis data of the 97 components with the highest VIP values per data, and incorporates all 125 data matrices into VMS as learning data (Fig. 2) built. ^R2 was 0.92.

(2-2) Each data has peak area values and yield values of the analysis data of the 97 components with the highest VIP values, and all 125 data matrices are randomly divided into two groups (for learning and for verification), Using one of the 63 matrix data for learning, a model (neural network) (Fig. 3) was constructed by VMS. ^R2 was 0.83. ^R2 in the verification data is 0.58, and it can be said that the model has good prediction performance.

(3) Building a model using the VIP value calculated from the model in (1) as an index (model using analysis data of two or more components)
(3-1) Model using analysis data of components with lower VIP values Analysis data of all components below VIP value 11th, Analysis data of all components below 21st, 3
A model (FIG. 4) was constructed by the OPLS method using the analysis data of all components below the 1st position and the analysis data of all the components below the 351st position.
A model that satisfies Q ² >0.50 is a model that uses analytical data of all components from 11th to 251st. Q ² >0.50 is not obtained even if all analytical data of components with VIP values of 261st and lower are used.

(3-2) Model using two pieces of analysis data of components with the top 10 VIP values Any combination of two of the analysis data of components with the top 1 to 10 VIP values (
45 patterns) were constructed by the OPLS method (Fig. 5).
Q ² >0.50 in any model.

(3-3) Model using analysis data of two consecutive components based on VIP value VIP value top 1st and 2nd, 11th and 12th, 21st and 22nd, … and 201st and 2
A model (FIG. 6) was constructed by the OPLS method using the analytical data of the 02-position component.
If the model is created using the analytical data of any two components selected from the top 30 VIP values, the value of Q ² is often Q ² >0.50.

(3-4) Model using analysis data of three consecutive components based on VIP value VIP value top 1st, 2nd and 3rd, 11th, 12th and 13th, 21st and 22nd 23rd place,
A model (Fig. 7) was constructed by the OPLS method using the analysis data of the components at positions 221, 222 and 223.
If the model is created using the analytical data of any three components selected from VIP values of 70 or higher, the value of Q ² is often Q ² >0.50.

(3-5) Model using analysis data of four consecutive components based on VIP value VIP value top 1st, 2nd, 3rd and 4th, 11th, 12th, 13th and 14th, 21st and 2nd place
A model (FIG. 8) was constructed by the OPLS method using analytical data for components at positions 2, 23, 24, .
If the model is created using the analysis data of any four components selected from the top 100 VIP values, the value of Q ² is often Q ² >0.50.

(3-6) Model using analysis data of five consecutive components based on VIP value VIP value top 1st to 5th, 11th to 15th, 21st to 25th, … and 251st ~2
A model (Fig. 9) was constructed by the OPLS method using the analytical data of the 55th component.
If the model is created using the analysis data of any five components selected from the top 100 VIP values, the value of Q ² is often Q ² >0.50.

(3-7) Model using analysis data of 6 consecutive components based on VIP value VIP value top 1st to 6th, 11th to 16th, 21st to 26th, … and 281st ~2
A model (Fig. 10) was constructed by the OPLS method using the analytical data of the 86th component.
If the model is created using the analysis data of arbitrary six components selected from VIP values 130 or higher, the value of Q ² is often Q ² >0.50.

(3-8) Model using analysis data of 7 consecutive components based on VIP value VIP value top 1st to 7th, 11th to 17th, 21st to 27th, … and 281st ~2
A model (FIG. 11) was constructed by the OPLS method using the analytical data of the 87th component.
If the model is created using the analysis data of any seven components selected from VIP values of 140 or higher, the value of Q ² is often Q ² >0.50.

(3-9) Model using analysis data of 8 consecutive components based on VIP value VIP value top 1st to 8th, 11th to 18th, 21st to 28th, … and 281st ~2
A model (FIG. 12) was constructed by the OPLS method using the analytical data of the 88th component.
If the model is created using the analysis data of arbitrary eight components selected from VIP values of 140 or higher, the value of Q ² is often Q ² >0.50.

(3-10) Model using analysis data of 9 consecutive components based on VIP value VIP value top 1st to 9th, 11th to 19th, 21st to 29th, … and 281st ~2
A model (FIG. 13) was constructed by the OPLS method using the analytical data of the 89th component.
If the model is created using the analytical data of 9 arbitrary components selected from the 140th or higher VIP value, the value of Q ² is often Q ² >0.50.

(3-11) Model using analysis data of 10 consecutive components based on VIP value VIP value top 1st to 10th, 11th to 20th, 21st to 30th, … and 281st ~
A model was constructed by the OPLS method using the analysis data of the 290th component (Fig. 14).
If the model is created using the analysis data of any 10 components selected from VIP values 160 or higher, the value of Q ² is often Q ² >0.50.

The number of components used for prediction is preferably as small as possible for easy prediction,
For example, it is 10 or less, preferably 5 or less, more preferably 3 or less, and most preferably 1. In order to improve the accuracy, the number of components is preferably large, for example, 11 or more, preferably 20 or more, more preferably 50 or more, more preferably 90
1 or more, most preferably 97. When prediction is performed using a small number of components, it is preferable to use components with higher VIP values or components with higher correlation coefficients for prediction.

(4) Building a yield prediction model using some component information selected from 431 In addition to using all 431 component peaks in Tables 3a to f, by using component peaks selected from them It is also possible to construct a highly accurate prediction model.
For example, among all 431 component peaks in Tables 3a to f, peak data of 301 components are selected in consideration of peak shape, average detection intensity between samples, etc., and this component peak is corrected as appropriate. Similarly, construct an OPLS model, calculate the VIP value of the constructed model (Tables 6a to 6d below), and construct a prediction model using the analysis data of the top 100 components (below). A highly accurate model with R ² =0.82 indicating predictability and Q ² =0.78 indicating predictability can be constructed (referred to as prediction model A in the examples below, FIG. 15).
<Top 100 VIP value component numbers in prediction model A>
7, 15, 17, 20, 21, 22, 23, 35, 37, 39, 42, 44, 51, 5
4, 57, 58, 68, 71, 73, 80, 85, 86, 90, 93, 95, 108, 1
16, 122, 131, 139, 149, 153, 157, 158, 160, 161, 1
65, 171, 176, 179, 187, 208, 214, 223, 227, 233, 2
37, 245, 252, 253, 261, 278, 279, 282, 283, 284, 2
94, 298, 299, 300, 304, 305, 308, 309, 310, 313, 3
16, 317, 318, 320, 325, 327, 328, 329, 330, 331, 3
52, 353, 355, 356, 357, 358, 359, 362, 363, 367, 3
80, 381, 385, 388, 389, 390, 392, 395, 396, 399, 4
21, 422, 428, 431.

Aspects and preferred embodiments of the present invention are presented below.
<1> A method for predicting soybean yield, comprising obtaining analysis data of one or more components from a leaf sample collected from soybean, and predicting soybean yield using the correlation between the data and soybean yield.
<2> Correct the analysis data of the one or more components by pooled QC method, <1>
The method described in .
<3> The method according to <1>, wherein the analysis data of the one or more components are corrected with an internal standard substance.
<4> The component has an accurate mass (m/z) of 139 to 1156 provided by mass spectrometry
The method according to any one of <1> to <3>, which is one or more selected from the following components.
<5> The component is one or more selected from the components listed in Tables 1a to 1c, defined by accurate mass (m/z) provided by mass spectrometry, <1> to <3> The method according to any one of
The <6> component is the component No. described in Tables 1a to 1c. 13, 14, 17, 20, 21,
22, 23, 28, 35, 36, 37, 39, 41, 42, 44, 47, 48, 51, 5
2, 54, 57, 58, 68, 71, 73, 80, 85, 86, 90, 91, 96, 98
, 99, 100, 107, 108, 110, 122, 125, 131, 134, 135,
137, 139, 142, 149, 150, 153, 157, 159, 160, 161,
171, 174, 176, 179, 181, 182, 188, 202, 208, 209,
214, 215, 217, 218, 228, 230, 235, 244, 245, 246,
247, 249, 251, 252, 253, 261, 264, 268, 275, 278,
279, 280, 282, 283, 284, 288, 294, 296, 298, 299,
305, 308, 310, 313, 317, 325, 327, 329, 330, 341,
347, 353, 355, 356, 363, 367, 369, 370, 384, 389,
<5> which is one or more selected from 395, 421, 422, 423, 428 and 431
The method described in .
The <7> component is the component No. described in Tables 1a to 1c. 14, 22, 23, 36, 37,
41, 42, 51, 52, 68, 90, 122, 139, 149, 159, 214, 22
8, 230, 235, 247, 249, 252, 253, 268, 275, 278, 28
4, 288, 298, 305, 308, 313, 317, 329, 347, 363, 39
5, 421, 422 and 428. The method according to <5>.
The <8> component is the component No. described in Tables 1a to 1c. 14, 23, 36, 37, 41,
51, 68, 90, 122, 149, 214, 230, 235, 247, 249, 252
, 275, 284, 298, 305, 308, 313, 317, 347, 363, 421
, 422 and 428. The method according to <5>.
The <9> component is _soyasaponin Bb; a monoglucoside body having a composition formula of C ₂₁ H ₂₂ O ₁₁ , in which glucose is bound to dehydroflavonol whose aglycone has a composition formula of C ₁₅ H ₁₂ O ₆ ; H ₄₀ O ₁₉ and the compositional formula of the aglycone is C ₁₅ H ₁₀ O _6
a triglycoside in _which one glucose and two rhamnose _are bound to _the flavonol _{of ;} The method according to <5>, which contains one or more selected from monoglucoside forms.
<10> A leaf sample is collected from the soybean from the primary leaf development stage to the seed enlargement stage, <1>
The method according to any one of ~<9>.
<11> A leaf sample is taken from the soybean from the primary leaf development stage to the flowering stage, <1>
The method according to any one of ~ <9>.
<12> The method according to any one of <1> to <11>, wherein the analysis data is mass spectrometry data.
<13><5> to <1, including the step of matching the analysis data of the components obtained from the leaf sample with the yield prediction model constructed using the analysis data of the components described in Tables 1a to 1c.
2>.
<14> The yield prediction model uses at least two of the top 10 VIP values calculated from the yield prediction model constructed using the component information listed in Tables 1a to 1c above.
13>.
<15> The yield prediction model uses at least two of the top 22 VIP values calculated from the yield prediction model constructed using the component information listed in Tables 1a to 1c above.
13>.
<16> The yield prediction model uses at least 3 of the top 63 VIP values calculated from the yield prediction model constructed using the component information listed in Tables 1a to 1c above, <
13>.
<17> The yield prediction model uses at least 4 of the top 94 VIP values calculated from the yield prediction model constructed using the component information listed in Tables 1a to 1c above.
13>.
<18> The yield prediction model uses at least 5 of the top 95 VIP values calculated from the yield prediction model constructed using the component information listed in Tables 1a to 1c above, <
13>.
<19> The yield prediction model uses at least 6 out of the top 126 VIP values calculated from the yield prediction model constructed using the component information listed in Tables 1a to 1c.
The method according to <13>.
<20> The yield prediction model uses at least 7 out of the top 137 VIP values calculated from the yield prediction model constructed using the component information listed in Tables 1a to 1c.
The method according to <13>.
<21> The yield prediction model uses at least 8 of the top 138 VIP values calculated from the yield prediction model constructed using the component information listed in Tables 1a to 1c.
The method according to <13>.
<22> The yield prediction model uses at least 9 of the top 139 VIP values calculated from the yield prediction model constructed using the component information listed in Tables 1a to 1c.
The method according to <13>.
<23> The yield prediction model uses at least 10 of the top 160 VIP values calculated from the yield prediction model constructed using the component information listed in Tables 1a to 1c above, <13> described method.
<24> The method according to any one of <14> to <23>, wherein the VIP value is calculated by a yield prediction model constructed using the ingredient information of all the ingredients listed in Tables 1a to 1c.
<25> A yield prediction model constructed using the analysis data of the components obtained from the leaf sample and the analysis data of the components listed in Tables A6a to A6d below selected from the components listed in Tables A1a to 1c above. The method according to <5>, including the step of collating.
<26> Including the step of collating the analysis data of the components obtained from the leaf sample with the yield prediction model constructed using the analysis data of the following 100 components selected from the components listed in Tables A1a to 1c above. , the method described in <5>,
Ingredient no. 7, 15, 17, 20, 21, 22, 23, 35, 37, 39, 42, 44
, 51, 54, 57, 58, 68, 71, 73, 80, 85, 86, 90, 93, 95,
108, 116, 122, 131, 139, 149, 153, 157, 158, 160,
161, 165, 171, 176, 179, 187, 208, 214, 223, 227,
233, 237, 245, 252, 253, 261, 278, 279, 282, 283,
284, 294, 298, 299, 300, 304, 305, 308, 309, 310,
313, 316, 317, 318, 320, 325, 327, 328, 329, 330,
331, 352, 353, 355, 356, 357, 358, 359, 362, 363,
367, 380, 381, 385, 388, 389, 390, 392, 395, 396,
399, 421, 422, 428, 431.
<27> Leaf samples are taken from soybeans from the stage of primary leaf development to the stage of seed enlargement, <25
> or the method according to <26>.
<28> Leaf samples are taken from the soybean from the stage of primary leaf development to the flowering stage, <25
> or the method according to <26>.
<29> The method according to any one of <25> to <28>, wherein the analytical data is mass spectrometry data.
<30> The yield prediction model is a model constructed using the OPLS method <13> ~ <
29>.
<31> The method according to any one of <13> to <29>, wherein the yield prediction model is a model constructed using machine learning/regression analysis techniques.
<32> Accurate mass is measured with an accuracy of 4 digits or more after the decimal point <4> to <3
1>.

1. Each cultivation test
The outdoor pot cultivation test data conducted from 2015 to 2017 will be described in detail. A total of 4 tests were performed.
1) 2015 pot trial (1):
Pot cultivation was carried out in Tochigi prefecture. Domestic field soil was used as the soil, and fertilization was performed so that nitrogen:phosphoric acid:potassium=3:6:6 (Kg/10a), and the soil was cultivated with a power tiller.
As the soil, this tilled soil was used. 1/2000a Wagner pots were used as pots, and 15 pots were prepared by packing about 8 L of the above soil per pot. June 25, 2015 3
Seeds were seeded at two locations in each pot by grain sowing (6 grains were used per pot). The cultivar used was 'Sato no Hohoemi'. At the stage of primary leaf development, one plant was thinned out at each site, and two plants were placed in each pot. Harvesting was carried out on November 9 (137 days after sowing). Note that 10 strains in 5 pots were used for yield prediction.

2) 2015 pot trial (2):
Pot cultivation was carried out in Wakayama prefecture. The soil uses domestic field soil, nitrogen: phosphoric acid:
Fertilization was performed so that potassium = 1: 6: 6, 3: 6: 6 and 10: 6: 6 (Kg / 10 a), and the soil after tillage was used (three types of fertilization conditions differing only in nitrogen content). set). 1/2000a Wagner pots were used as pots, and about 8 L of the above soil was packed in each pot, and 15 pots in total of 45 pots were prepared for each fertilization condition. On July 1, 2015, 3 grains were sown at 2 locations in each pot (6 grains were used in each pot). "Fukuyutaka" was used as the variety.
At the stage of primary leaf development, one plant was thinned out at each site, and two plants were placed in each pot. Harvest was performed on November 11 (133 days after sowing). It was planned to use 10 strains in each of the 5 pots for yield prediction, but a total of 29 strains were used because one strain was missing.

3) 2016 pot test:
Pot cultivation was carried out in Tochigi Prefecture. The test was conducted using field soil from Japan as the soil. 1/2000a Wagner pots were used as pots, and 75 pots were prepared by packing about 8 L of the above soil per pot. After standing for several days, seeding was performed in the same manner as in 2015. Sowing is 2
It was done on July 1, 016 and harvested on November 15th. In addition, the variety is "Sato no Hohoemi"
was used. For yield prediction, 46 strains in 23 pots were used.

4) 2017 pot test:
Pot cultivation was carried out in Tochigi prefecture. The soil used was soil from a domestic field, and 125 g/m ² of magnesium lime (Kyowa) and 100 g/m ² of phosphate-containing oil cake (Daiei Bussan) were added as fertilizers.
Added soil (1 ×) and soil mixed with half the amount of soil with added fertilizer and soil with no added fertilizer (0
. 5×) were used. As for the varieties, four varieties of "Sato no Hohoemi", "Fukuyutaka", "Enrei" and "Yuagari Musume" were used. A 1/2000a Wagner pot was used as the pot, and about 8 L of the above soil was packed in each pot. After being allowed to stand for several days, 4 grains of each variety were sown at 2 locations in each pot (8 grains were used in each pot). At the stage of primary leaf development, one plant was thinned out at each site, and two plants were placed in each pot. Seeding was performed on July 4, 2017, and harvesting was performed sequentially from October onwards, starting from strains that had reached maturity and were judged to be suitable for harvesting.

2. leaf sampling
Sampling of leaves was carried out in the daytime on days 28-32 after seeding in each cultivation test (generally 10:00-15:00). The growth stage of the soybean at this time slightly differed depending on the year, cultivation conditions and variety, but the leaf age was generally about 5-7. The leaf age referred to here is 1
The value was obtained by counting the number of sheets of true leaves that were developed at the top from the bottom. Sampling of leaves is performed on three compound leaves that are 1 or 2 leaves older than the topmost true leaf,
A central compound leaf was collected. However, when the central compound leaf was severely damaged such as by insect damage, another compound leaf was collected. The harvested leaves were wrapped in aluminum foil and immediately frozen in liquid nitrogen.
Stopped the metabolic reaction. Frozen samples were brought to the laboratory in a frozen state and lyophilized to dryness. This dried sample was subjected to the extraction procedure described below. In addition, 2
In the pot cultivation tests in 2015 and 2016, each individual was collected, and the yield data of the corresponding individual was used. On the other hand, in the 2017 pot cultivation test, sampling was performed collectively for each pot, that is, every two individuals, and the average value of the two individuals was used as the yield data.
In addition, the number of days after sowing when sampling leaves is as follows.
* 2015 pot test (1): July 25, 2015 (30 days after sowing)
* 2015 pot test (2): July 29, 2015 (28 days after sowing)
* 2016 pot test: August 2, 2016 (32 days after sowing)
* 2017 pot test: August 3, 2017 (30 days after sowing)

3. Measurement of final grain yield
Whole grains were collected from each plant after the cultivation test and dried at 80° C. for 2 to 3 days. This dry weight (gDW/individual) was used for the yield data. As described in 2-2, the average data of two individuals (per pot) in the 2017 test was counted as one, and the pot test data from 2015 to 2017 totaled 125. Yield data was a minimum of 0.9 gDW/plant and a maximum of 42.5 gDW/plant as shown in Tables 2a-2c.

4. Extraction of components from harvested leaves
Freeze-dried leaf samples were manually ground with a spatula as much as possible.
After grinding, 10 mg was weighed into a 2 mL tube (Safe Lock Tube, Eppendorf), one 5 mm diameter zirconia ball was added to the tube, and a bead grinder (MM40
0, Retsch) at 25 Hz for 1 minute. The extraction solvent was 500 ng/mL of lidocaine (Wako Pure Chemical Industries, #120-02671) as an internal standard.
A 0 v/v % methanol aqueous solution was used. 1 mL of the prepared extraction solvent was added to the tube after pulverization, and homogenized extraction was performed at 20 Hz for 5 minutes using the same bead pulverizer. After extraction,
Centrifuge for about 30 seconds in a tabletop centrifuge (Cibitan) at about 2,000 × g, and pass through a 0.45 μm hydrophilic PTFE filter (DISMIC-13HP 0.45 μm syringe f
ilter, ADVANTEC) to obtain an analytical sample.

5. Analysis of leaf samples by LC/MS
Analysis of leaf extract samples was performed on an Agilent HPLC system (Infinity1
260 series) on the front and AB SCIEX Q-TOFMS equipment (Trip
LC/MS analysis was performed using leTOF4600) as a detector. As a separation column in HPLC, a core-shell column Capcell core C18 manufactured by Shiseido Co., Ltd.
(2.1 mm ID x 100 mm, particle count 2.7 μm) and a guard column (2.1 m
m I. D. x5 mm, particle size 2.7 μm) was used and the column temperature was set to 40°C.
The autosampler was kept at 5°C during the analysis. 5 μL of analysis sample was injected. As eluents, A: 0.1 v/v % formic acid aqueous solution and B: 0.1 v/v % formic acid acetonitrile solution were used. Gradient elution conditions were held at 1 v/v % B (99 v/v % A) from 0 min to 0.1 min and 1 v/v % B to 99.5 v/v from 0.1 min to 13 min. %B was increased and held at 99.5 v/v %B from 13.01 min to 16 min. Flow rate is 0.5 mL
/min.

As for the conditions of the mass spectrometer, the ionization mode was positive mode, and ESI was used as the ionization method. In this analysis system, the eluted ions are scanned by TOFMS for 0.1 second,
While repeating the cycle of selecting 10 ions with high intensity among them and subjecting each to MS/MS for 0.05 seconds, molecular ion information (accurate mass, m/z) by TOFMS scanning and MS/MS scanning Structural information was obtained from the resulting fragments. Mass measurement range is m/z 100-1,250 for TOFMS and m/z for MS/MS
It was set at 50-1,250. The parameters for each scan are GS1=50, GS2=50, CUR=25, TEM=450, ISVF=5500,
Set DP = 80 and CE = 10, for MS/MS scans GS1 = 50, G
S2=50, CUR=25, TEM=450, ISVF=5500, DP=80, CE=
30, CES=15, IRD=30 and IRW=15.

6. Create Data Matrix
Data processing was performed as follows. First, the MarkerView ^™ Software
Peak extraction was performed using (AB SCIEX). Peak extraction conditions (“peak f
indexing option”) is a peak corresponding to a retention time of 0.5 minutes to 16 minutes,
Subtraction in the item of "Enhance Peak Finding"
20 scans of offset, Minimum spectral peak width
th is 5 ppm, Subtraction multi. Factor is 1.2, Mi
10 scans of nimum RT peak width, Noise threshold
Set ld to 5, Assign charge state in the "More" item
was checked. As a result, 12,444 peak information was obtained.

Next, an alignment process was performed to align the detected peaks between each analyzed sample. Alignment processing conditions (“Alignment & Filtering”
) is the Retention time tole in the "Alignment" item
The lance was set to 0.20 minutes and the mass tolerance to 10.0 ppm. In addition, the Intensity threshold in the item of "Filtering"
10, check Retention time filtering, Rem
With ove peaks in <3 samples, Maximum number of p
eaks was set to 50,000. The retention time was corrected using the peak of lidocaine in the item of "Internal standards".

Next, removal of isotopic peaks was performed. Isotopic peaks are automatically recognized by the software at the time of peak extraction and are labeled as "isotopic" in the peak list.
Relevant peaks were deleted by sorting by "isotopic". As a result, the peak is 10,1
reduced to 12 peaks.

Next, peak intensity correction between samples was performed. In this analysis, in addition to the sample, a sample called pooled QC was prepared by mixing a constant amount from all the samples, and the pooled QC analysis was performed once every nine times. From these full QC analysis results, "
Assuming that QC samples were analyzed when each sample was analyzed, what would happen to each peak intensity? Sensitivity correction was performed between For this processing, free software (LOWESS-Normalization-Tool) provided by RIKEN was used. Finally, the relative standard deviation (RSD) of 10,112 peaks was calculated using the measured 30 QC run data, peaks with high variability with RSD>30% were removed, and finally 431 peak data , that is, analytical data of 431 components were obtained. The analytical data obtained are presented in Tables 3a-3f. Subsequent analysis was performed using these data.

7. Correlation analysis
Analysis data and corresponding yield data for 431 components in 125 leaves obtained, namely 12
Correlation analysis was performed using 5×432 matrix data. The single correlation coefficient r between the analysis data and the yield data of each component and the p-value were calculated by the test of non-correlation. The results are shown in Tables 4a-4f. "Component No." in the table is for the sake of convenience, in which the 431 components are numbered from the smallest mass number when arranged in order of mass. In addition, the analysis results include retention time information as well as mass information, but according to JP-A-2016-57219, if an accurate mass number with four or more decimal places is used, a plurality of data can be obtained regardless of the retention time. It has been shown that mass spectrometry data can be compared and analyzed between mass spectrometry samples. Therefore, retention time information was removed and only accurate mass information was described.

The results obtained in correlation analysis indicated that components with constant correlation coefficients were significantly correlated with yield. There are 118 components with the absolute value of the correlation coefficient |r|>0.51, |r|>0.66
It was found that there are 28 components that result in

8. Model construction/evaluation
A multivariate analysis method is used to construct a yield prediction model using analytical data of two or more components, and SIMCA ver. 14 (Umetrics) was used. For the predictive model, regression analysis was performed using the peak area values of the corrected analytical data of the components having each accurate mass as the explanatory variables and the yield values as the objective variables. Regression analysis was performed by the OPLS method, which is an improved version of the PLS method.

The prediction model evaluation method is mainly judged by two indices. One is R ² representing prediction accuracy,
The other is ^Q2 , which represents predictability. ^R2 is the square of the correlation coefficient between the measured values of the data used to construct the prediction model and the predicted values calculated by the model, and the closer to 1, the higher the prediction accuracy. On the other hand, ^Q2 is the result of the above cross-validation, and represents the square of the correlation coefficient between the measured value and the predicted value, which is the result of repeated model verification. From a prediction point of view, if at least Q ² >0.50, the model is said to have good predictive power (Triba, MN et al., Mol. BioSyst. 2015, 11, 13-19 .), Q ² >0.50
was used as the criterion for model evaluation. Since R ² >Q ² is always satisfied, Q ² >0.50 simultaneously satisfies R ² >0.50.

8-1. Construction and evaluation of models using all data of 431 components
An OPLS model was constructed to predict yield from a total of 125 data matrices, with analytical data peak area values and yield values for 431 components per datum. At the time of construction, the peak area value and yield data of analytical data for each component were converted to mean 0 and variance 1 by autoscaling. As a result of model construction, R ² = 0.87, which indicates prediction accuracy, and Q ² =, which indicates predictability
was 0.78. The results are shown in FIG. By using this prediction model, it was shown that a model with high prediction performance can be constructed by using the component composition contained in leaves for about one month of cultivation, and early yield prediction is possible.

8-2. Calculation of VIP value
VIP (Variable Importance in
A contribution to the model performance given to each component called the Projection (Variable Importance in Projection) value is given. The larger the VIP value, the greater the degree of contribution to the model, and the absolute value of the correlation coefficient is also correlated. A list of VIP values is shown in Tables 5a-5f.

8-3. machine learning model
Prediction model construction is not limited to the OPLS method, and can be constructed by various methods. As another example, we constructed a prediction model using machine learning. Machine learning is one of the research subjects in artificial intelligence, that is, AI, and is currently being applied to various fields.
Here, a model was constructed by machine learning using the analysis data of the top 97 components of the VIP value calculated from the model using all the data constructed in 8-1 above. The analysis tool includes Visual Mining Studio (hereinafter referred to as VMS, N Co., Ltd.).
TT Data Mathematical System) was used.

8-3-1. Model construction using all 125 data Each data had peak area values and yield values of analysis data of 97 components with the highest VIP values, and all 125 data matrices were taken into VMS and used as learning data. For model construction, the Model Optimizer function was used to search for the optimum model from four types of decision tree, random forest, neural network, and support vector machine. Model calculation optimizes the parameters of each model, and cross-validation is also performed to construct a model that does not cause overfitting. As a result, random forest was selected as the model. The square of the correlation between the actual value and the predicted value (R ² ) was 0.92, and a highly accurate model was constructed. The results are shown in FIG.

8-3-2. Model construction using half of the data and prediction verification with the remaining half Each data has the peak area value and yield value of the analysis data of the 97 components with the highest VIP value, and all 125 data matrices are randomly divided into two groups. It was divided, a model was constructed by VMS using 63 matrix data on one side, and prediction verification was performed on the remaining 62 data. When building a model with Model Optimizer as in 8-3-1,
A neural network was selected as the model. The square of the correlation between the measured values and the predicted values in the 63 data used in the model is 0.83, and the square of the correlation between the measured values and the predicted values in the 62 data not used in the model is 0.58. Met. Although the accuracy of the predicted value with the validation data was lower than that with the training data, it was shown that a certain level of prediction was possible. The results are shown in FIG.

8-4. Model construction using VIP value as an index (model using analysis data of two or more components)
ru)
VIP value ranking, which is the contribution of each component to the model constructed in 8-1 (Tables 4a-4
A model was constructed with multiple components based on f). For convenience, Q ² >0.50 was used as a criterion for model performance, although it is not particularly limited.

8-4-1. Model using analysis data of components with lower VIP values Analysis data of all components below VIP value 11th, analysis data of all components below 21st, 3
An OPLS model was constructed using the analysis data of all the components below the 1st position and the analysis data of all the components below the 351st position. As a result, the model that satisfies Q ² > 0.5 is the model using the analysis data of all the components below 11th to the analysis data of all the components below 251st, and all the analysis data of the components below VIP value 261st. (Fig ^. 4).

8-4-2. Model using two analysis data of the top 10 VIP value components Combination of any two of the analysis data of the top 1 to 10 VIP value components (
45 patterns), an OPLS model was constructed. As a result, it was found that Q ² >0.50 was satisfied in any model. From this, it was shown that if two metabolites with the top 10 VIP values are included, a model with a certain degree of predictiveness can be constructed (Fig. 5).
.

8-4-3. Model using analysis data of two consecutive components based on VIP value VIP value top 1st and 2nd, 11th and 12th, 21st and 22nd, ... and 201st and 2nd
Each OPLS model was constructed using the analysis data of the 02nd component. resulting in,
For the first time, the model did not satisfy Q ² >0.50 when using analytical data for two components at positions 31 and 32. After that, ^Q2 tended to decrease. From this, if the VIP value is about 30th or higher, the analysis data of any two components from among them can be used to generally satisfy the ^Q2 criteria, but the analysis data 2 of the component with a VIP value of about 30th It was shown that the criteria were not satisfied with only one (Fig. 6).

8-4-4. Model using analysis data of three consecutive components based on VIP value VIP value top 1st, 2nd and 3rd, 11th, 12th and 13th, 21st, 22nd and 23rd,
... and OPLS using the analytical data of the components at positions 221, 222 and 223, respectively
I built the model. As a result, Q ² >0.50 was not satisfied for the first time in the model using the analytical data of the three components at positions 71, 72 and 73. After that, ^Q2 tended to decrease. From this, if the VIP value is about 70th or higher, the analysis data of any 3 components from among them can be used to generally meet the criteria of ^Q2 . It was suggested that the criteria were not satisfied with only one (Fig. 7).

8-4-5. Model using analysis data of four consecutive components based on VIP value VIP value top 1st, 2nd, 3rd and 4th, 11th, 12th, 13th and 14th, 21st and 2nd
OPLS models were constructed using the analysis data of the components of 2nd, 23rd and 24th, . As a result, 101st and 102nd
Q ² > 0 for the first time in the model when using the analytical data of four components at positions 103 and 104.
. did not reach 50. After that, ^Q2 tended to decrease. From this, VIP
If the value is about 100 or more, Q
Although the criteria of ² are generally satisfied, it was suggested that the criteria were not satisfied with only 4 pieces of analysis data for components with a VIP value of about 100th or lower (Fig. 8).

8-4-6. Model using analysis data of five consecutive components based on VIP value VIP value top 1st to 5th, 11th to 15th, 21st to 25th, … and 251st to 2
An OPLS model was constructed using the analysis data of the 55th component. resulting in,
Q ² > 0.5 for the first time in the model when using the analytical data of 5 components from 101st to 105th
did not meet 0. After that, ^Q2 tended to decrease. From this, VIP value 1
If it is about 00th or higher, the criteria for ^Q2 is generally satisfied by using analysis data of any 5 components from among them, but only 5 analysis data of components with a VIP value of about 100th or less fulfills the criteria. (Fig. 9).

8-4-7. Model using analysis data of 6 consecutive components based on VIP value VIP value top 1st to 6th, 11th to 16th, 21st to 26th, … and 281st to 2
An OPLS model was constructed using the analytical data of the 86th component. resulting in,
Q ² > 0.5 for the first time in the model when using analytical data of six components from 131st to 136th
did not meet 0. After that, ^Q2 tended to decrease. From this, VIP value 1
If it is about 30th or higher, the criteria for ^Q2 can be roughly met by using analysis data for any 6 components from among them, but only 6 analysis data for components with a VIP value of about 130th or lower will meet the criteria. (Fig. 10).

8-4-8. Model using analysis data of 7 consecutive components based on VIP value VIP value top 1st to 7th, 11th to 17th, 21st to 27th, … and 281st to 2
Each OPLS model was constructed using the analysis data of the 87th component. resulting in,
Q ² > 0.5 for the first time in the model when using the analytical data of seven components from 141st to 147th
did not meet 0. After that, ^Q2 tended to decrease. From this, VIP value 1
If it is about 40th or higher, the criteria for ^Q2 is generally met by using analysis data for any 7 components from among them, but only 7 analysis data for components with a VIP value of about 140th or lower satisfies the criteria. (Fig. 11).

8-4-9. Model using analysis data of 8 consecutive components based on VIP value VIP value top 1st to 8th, 11th to 18th, 21st to 28th, … and 281st to 2
Each OPLS model was constructed using the analysis data of the 88th component. resulting in,
Q ² > 0.5 for the first time in the model when using analytical data of eight components from 141st to 148th
did not meet 0. After that, ^Q2 tended to decrease. From this, VIP value 1
If it is about 40th or higher, the criteria for ^Q2 is generally satisfied by using analysis data of any 8 components from among them, but only 8 analysis data of components with a VIP value of about 140th or lower fulfills the criteria. (Fig. 12).

8-4-10. Model using analysis data of 9 consecutive components based on VIP value VIP value top 1st to 9th, 11th to 19th, 21st to 29th, … and 281st to 2
Each OPLS model was constructed using the analysis data of the 89th component. resulting in,
Q ² > 0.5 for the first time in the model when using analytical data of nine components from 141st to 149th
did not meet 0. After that, ^Q2 tended to decrease. From this, VIP value 1
If it is about 40th or higher, the analysis data of 9 components from among them will be used, and the criteria for ^Q2 will be mostly met, but the criteria will not be met with only 9 analysis data of components with a VIP value of about 140th or lower. was suggested (Fig. 13).

8-4-11. Model using analysis data of 10 consecutive components based on VIP value VIP value top 1st to 10th, 11th to 20th, 21st to 30th, ... and 281st
An OPLS model was constructed using the analysis data of the 290th component. As a result, Q ² > 0 for the first time in the model when using the analytical data of 10 components from 161st to 170th
. did not reach 50. After that, ^Q2 tended to decrease. From this, VIP
If the value is about 160th or higher, the criteria for ^Q2 can be roughly met by using analysis data of any 10 components from among them, but if only 9 analysis data of components with a VIP value of about 140th or less are used, the criteria are not met. It was suggested that it does not satisfy (Fig. 14).

8-5. Prediction model construction and evaluation using analysis data of 100 components
Of the 431 component peaks in Tables 3a to 3f, 301 peak data were selected in consideration of peak shapes, average detection intensity between samples, and the like. For these 301 component peaks, each peak area relative value to the peak area of lidocaine added as an internal standard was calculated and corrected instead of correcting the peak intensity by the above-described pooled QC. Using the corrected data, the analysis tool SIMCA described above was used to construct a model in the same manner as the method described in 8-1 above. That is, an OPLS model for predicting yield was constructed from a total of 125 data matrices, with peak area values and yield values of analytical data for 301 components per data. Calculate the VIP value of the constructed model (list of VIP values in Tables 6a-6d
shown in ), when a model was further constructed using the analysis data of the top 100 components, a highly accurate model was constructed with R ² = 0.82 indicating prediction accuracy and Q ² = 0.78 indicating predictability ( hereinafter referred to as "prediction model A"). The results are shown in FIG.

9. Selecting the best fertilizer and materials for each field
9-1. Material selection method
Soybean production fields to be evaluated on March 8, 2019 (paddy field conversion field, rice - barley -
Soil was collected from soybean crops (three crops in two years). To 10 L of the field soil collected on March 18, 1 g of Best Match (600 for exclusive use of Yugari Musume, Kaneko Seed Co., Ltd.) was added as a common base fertilizer (
Equivalent to 20 kg/10a at a depth of 20 cm of mixed soil, the same amount as applied to the field).
There, set different test plots 1 to 10 including the following 6 types of materials alone or their combinations that are candidates for selection, add the recommended amount of each material and mix well, then 1 L each in 8 poly pots filled. Soybean seeds (variety: Sato no Hohoemi, 2018 product from Yamagata Prefecture,
Seed size 8.5 mm or more) is sown, and two pots are placed in one bat,
Cultivation was started on an indoor cultivation rack (under fluorescent lamps). Water was appropriately supplied to all vats under the same conditions.
Four weeks after seeding, on April 17, leaf sampling was carried out in the same manner as described in 2 above. Leaves sampled from two individual bats were mixed to obtain one sample, and the extraction and analysis described in 4 and 5 above were performed to obtain component data of each sample. In addition, the following was used for the candidate of the material to select. Details of the test plots are shown in Table 7.

・MIX compost (Kawaguchi Fertilizer Co., Ltd.)
・Natural saponin lees (Saitama Noko Ikuka Sales Co., Ltd.)
・Soybean cake (Nissin OilliO Group Inc.)
・Sand-like yorin (Akagi Bussan Co., Ltd.)
・Microelement 8 (Aminol Chemical Laboratory)
・Ammonium sulfate (Akagi gardening)

FIG. 16 shows the result of yield prediction by incorporating the acquired component data into the model A described above. In addition, FIG. 16 shows the predicted yield difference from test plot 1, which is the control plot, by using the data of each test plot n=4 as an average value±standard deviation. In test plots 2 (single application) and 9 (mixture with other agents containing MIX compost) in which MIX compost was applied, a higher yield than the control plot was predicted. Therefore, the MIX compost common to test plots 2 and 9 was selected as a material to be used in field cultivation.

9-2. Confirmation of predicted values in field tests and actual measurement yield results
A field yield evaluation test of the selected material (MIX compost) was carried out in the same year in a field planned for soybean production, which is the object of evaluation. Before sowing on June 25, the sowing date, three sites in the field (three replicates)
material was applied to In addition, 3 repetitions of non-application plots were set as comparison controls. The test plot area per site was set to 2 m ² , and the same amount of material as in the indoor cultivation test was applied according to the area. On July 26, one month after seeding, leaves were sampled and yield prediction was performed in the same manner as in the selection test to examine whether the yield predictability in the field test was the same as that in the room. For leaf sampling, 2 samples were obtained by mixing 5 individuals in each repetition, and a total of 6 samples were obtained in 3 repetitions. The results are shown in FIG. In addition, FIG. 17 shows the average value±standard deviation of the predicted yield using the data of n=6 in the unapplied area and the MIX compost area.
The predicted yield in the field test was significantly higher (p<0.05, Student's t-test) in the MIX compost application area than in the non-application area, confirming the same results as in the selection test.

Cultivation was continued after leaf sampling, and harvesting was carried out on November 12th. Yield data was obtained by the method described in 3 above. The yield value of the unapplied plot and the MIX compost plot is 1 from within each replicate.
It was calculated from the values of 26 strains excluding 2 strains with the highest yield and 2 strains with the lowest yield among all 30 strains from which 0 individuals were harvested.
The results are shown in FIG. In addition, FIG. 18 shows the average value±standard deviation of the measured yield using the data of n=26 in the unapplied area and the MIX compost area.
While it was 18.7 gDW / strain in the unapplied section, it was 20.7 gDW in the MIX compost section.
/ strain, and a 10.7% increase in revenue was observed.
From the above results, it is shown that by using this yield prediction model, it is possible to select the materials suitable for the field in the short period before field cultivation, and to increase the yield in field cultivation by applying the selected materials. was done.

10. Examination of whether yield prediction is possible at 2 weeks and 8 weeks after sowing
In the above, the yield prediction was performed about one month after seeding. The yield predictability at earlier (2 weeks after sowing) or later (8 weeks after sowing) was examined below.
This test was carried out from March 2019 by pot cultivation in a glass house in Kanagawa Prefecture.
Thirty 1/5000a Wagner pots were used as the pots, and 4 L of soil was filled in each pot using domestic field soil. All 30 pots were divided into 3 groups of 10 pots each of 3 test plots A, B and C with different fertilizer application amounts, and cultivation was carried out. Magump K (Hyponex) was used as fertilizer, 10 g/pot in test area A, 5 g/pot in test area B, and 20 g in test area C.
/ Pots were applied before sowing. Enlays were used as soybean seeds and sown on March 14th.
In addition, two seeds were sown in one pot and thinned out so as to obtain one plant/pot at the initial leaf development stage. On April 1, 2 weeks after seeding and on May 9, 8 weeks after seeding, leaf sampling was performed in the same manner as described in 2 above, and extraction and analysis described in 4 and 5 above. was performed to obtain component data for each sample. The final harvest was performed on June 27, and the yield data for each strain was obtained by the method described in 3 above. In addition, water was appropriately applied after sowing until harvest.

The predicted yield of each test plot calculated using model A from leaves sampled two weeks and eight weeks after sowing was compared with the actual yield measured at the time of harvest. The results are shown in FIG. In addition, Fig. 1
9 shows the mean ± standard deviation of predicted yield and measured yield using n = 10 data of test plots A to C.
The predicted yield at 2 weeks after sowing reflects the difference in the measured yield of each test section, and it was shown that the yield property can be evaluated even at 2 weeks after sowing. On the other hand, 8 weeks after sowing did not reflect the difference in the measured yield of each test plot even though it was closer to the harvest time.
From these results, it was shown that the present yield prediction method is characterized by the ability to predict with higher accuracy in the early stage of growth, such as 2 to 4 weeks after sowing.

11. Yield prediction model construction using field data
11-1. Overview of each field trial
11-1-1. Field test in 2015 Cultivation was carried out in the producer's field in Tochigi Prefecture (paddy field conversion field, rice - barley - soybean 2 years 3 crops). Fertilization before seeding was carried out so that nitrogen:phosphate:potassium=2.4:8:8 (Kg/10a), and 40 kg of lime silicate fertilizer was added per 10a. Sato no Hohoemi was used as the cultivar. Seeded on June 15, 2015. Leaves were sampled as described below and harvested on November 1 (138 days after sowing). For yield prediction, a total of 29 individuals of 9 individuals or 10 individuals were collected from 3 points in the field.

11-1-2. Field test in 2016 (1)
Cultivation was carried out in Miyagi prefecture. Fertilization before sowing was nitrogen:phosphate:potassium=1.5:1.
5:1.5 (Kg/10a). The cultivar used was Enrei. 201
Seeds were sown on June 10, 6. Leaves were sampled as described below and harvested in mid-November. Based on the appearance of the soil, the field was divided into two sections, A and B, and a total of 24 individuals, 12 from each of sections A and B, were collected for yield prediction.

11-1-2. Field test in 2016 (2)
Cultivation was carried out in the producer's field in Tochigi Prefecture (paddy field conversion field, paddy rice - barley - soybean 3 crops in 2 years, but different field from the 2015 field test). Fertilization before sowing was performed in the same manner as in 2015. Sato no Hohoemi was used as the cultivar. Seeded on June 7, 2016. Leaves were sampled as described below and harvested in late November (approximately 160-170 days after sowing). In addition, it was planned to collect a total of 30 individuals, 10 individuals each from 3 points in the field, for yield prediction.
In the field in 2016, a large number of green spots occurred, and 8 individuals (out of 30 individuals) could be harvested for yield prediction.

11-1-4. Field test in 2017 Cultivation was carried out in three fields (T, YS, YM) of producers in Tochigi Prefecture. Fertilization before seeding was carried out as usual. Sato no Hohoemi was used as the cultivar. Seeding was carried out in field T on June 27, 2017, in field YM on June 29, and in field YS on July 7. Leaves were sampled as described later.
It was carried out on January 2nd and in field YS on November 2nd. For yield prediction, 5 individuals were collected from 5 locations in the field, and the 5 individuals were collectively used as one sample. That is, a total of 15 samples (for 75 individuals) of 5 samples (for 25 individuals) from each field were collected. In addition, the cropping system of each field is as follows: Field T has three crops in two years of paddy rice - wheat - soybeans, and field YM has three crops of rice - paddy rice - wheat - soybeans
There were 4 crops a year, and paddy rice was monocropped for more than 10 years in the field YS.

11-2. leaf sampling
Leaves were sampled in the same manner as described in 2 above. The number of days from sowing and the schedule for sampling in each field test are as follows.
* 2015 field test: July 15, 2015 (30 days after sowing)
* 2016 field test (1): July 21, 2016 (41 days after sowing)
* 2016 field test (2): July 6, 2016 (29 days after sowing)
* 2017 Field Test_Field T: July 28, 2017 (31 days after sowing)
* 2017 Field Test_Field YM: July 31, 2017 (32 days after sowing)
* 2017 Field Test_Field YS: August 7, 2017 (31 days after sowing)

11-3. Predictive model building
A total of 76 leaf samples obtained in the field tests conducted in 2015-2017 were extracted and analyzed as described in 4 and 5 above to obtain analytical data for each sample. These data were analyzed as described in 6 above, and 43
Analytical data for one component was obtained. The measured yield values used to construct the prediction model are for each field (however, 2
016 field test (1) was divided into section A and section B), and the average value was used. As shown in Table 8, the average yield value of each field was 10.27 gDW/plant at minimum and 27.66 gDW/plant at maximum.

Based on these results, the OPLS model that predicts the yield of each field was applied using the aforementioned analysis tool SI
Built using MCA. As a result of model building, R ² =0.86 indicating prediction accuracy and Q ² =0.76 indicating predictability. The results are shown in FIG.
Using this prediction model, it was shown that by using the component composition contained in leaves for about one month after cultivation, a model with high prediction performance can be constructed even for samples collected in the field, and early yield prediction is possible. .

Claims (4)

1. A soybean yield prediction method comprising obtaining analysis data of one or more components from a leaf sample collected from soybean, and predicting the soybean yield using the correlation between the data and the soybean yield, A method, wherein the above component is soyasaponin Bb. 2. The method of claim 1, wherein the leaf sample is taken from the soybean from the stage of primary leaf development to the stage of grain enlargement. 3. The method of claim 1 or 2, wherein the leaf sample is taken from the soybean from the primary leaf development stage to the flowering stage. The method according to any one of claims 1 to 3, wherein the analytical data are mass spectrometric data.

Images