WO2024248014A1 - Method for sorting seeds - Google Patents

Abstract

The present invention provides a method for sorting plant seeds having a desired trait, the method including the following first to seventh steps: (First step) A step for extracting some seeds from a seed group. (Second step) A step for constructing a data set with respect to the extracted seeds, the step for constructing a data set including the following (a) to (c): (a) beaming light onto the extracted seeds and acquiring spectral data and/or spectral data derived therefrom; (b) subjecting the seeds for which spectral data has been acquired to quality inspection with respect to the desired trait, and determining the advance quality rate of the extracted seeds, the advance quality rate being defined by the following formula: [Advance quality rate (%)] = [number of seeds having desired trait from among seeds subjected to quality evaluation] / [number of seeds subjected to quality evaluation] × 100; (c) associating the spectral data for each seed and the result of the quality inspection. (Third step) A step for applying multivariate discriminant analysis to the data set constructed in the second step and thereby deriving candidates for a good/poor discrimination model for calculating a discrimination score for the desired trait in each seed. (Fourth step) A step for selecting a discrimination model from the candidates for a good/poor discrimination model derived in the third step, the step including selecting as a good/poor discrimination model a candidate that satisfies the criterion that the conformity rate exceeds the advance quality rate under a standard sorting condition, where the recovery rate and the conformity rate are defined by the following formulae, and a sorting condition that the advance quality rate and the recovery rate have the same value is defined as the "standard sorting condition": [Recovery rate (%)] = [Number of seeds selected on the basis of discrimination score from among seeds subjected to quality evaluation] / [number of seeds subjected to quality evaluation] × 100 [Conformity rate (%)] = [Number of seeds with desired trait from among seeds subjected to quality evaluation and selected on the basis of discrimination score] / [number of seeds selected on the basis of discrimination score from among seeds subjected to quality evaluation] × 100 (Fifth step) A step for beaming light onto the remaining seeds not extracted from the seed group in the first step and/or other seeds obtained under essentially the same conditions as those for said seeds and acquiring spectral data, and applying the discrimination model selected in the fourth step to the spectral data to determine the discrimination score for each seed. (Sixth step) A step for determining a threshold for the discrimination score on the basis of the discrimination scores determined in the fifth step. (Seventh step) A step for comparing the discrimination scores of the respective seeds determined in the fifth step with the threshold for the discrimination score determined in the sixth step, and thereby recovering seeds predicted to have the desired trait.

Description

How to Select Seeds

The present invention relates to a method for selecting seeds, and more specifically, to a method for selecting plant seeds having desired characteristics by analyzing the spectral data of the seeds using multivariate statistical analysis.

The world population continues to grow, and is expected to reach 8 billion in 2022, 8.5 billion in 2030, and 9.7 billion in 2050. Therefore, the importance of stable food production and supply is increasing in order to ensure the survival of human society. One effective way to solve this problem is to develop high-yielding varieties and spread them around the world. However, no matter what variety is used to produce crops, the excellent characteristics of that variety cannot be demonstrated without good-quality seedlings. In ancient Japan, there was a phrase "naehansaku," which means "half of the growth is determined by the seedlings." Going further, good-quality seedlings cannot be produced without good-quality seeds. Therefore, how to produce and prepare high-quality seeds, how to preserve them without deteriorating their quality, and how to evaluate the current state of their quality are urgent technical issues that humanity must work together to address.

The importance of seed quality control is strongly recognized by public institutions both in Japan and abroad. The International Seed Testing Association (ISTA) provides guidelines on quality testing methods for almost all seeds of major crop species. The Japanese government also enforces the Seed and Seedlings Act, which sets quality standards for seeds for agricultural production distributed on the market. One of the reasons for such public monitoring is that seed quality (especially germination and genetic purity) is usually very difficult to identify by the appearance of the seed or other simple criteria. If low-quality seeds are accidentally spread, it may not only lead to local food shortages but also to poverty for farmers.

Although official quality control is certainly useful in preventing seed-related crop failures, there are always two sides to everything. For example, if the quality rate that seeds must meet is set at 90% from the perspective of official quality control, a seed lot with a quality rate just 1% lower than 90% (i.e., an 89% quality rate) will lose its commercial value. However, in reality, due to the fierce quality competition among seed suppliers, even such legal standards have become irrelevant or uncompetitive. As a result, a large amount of seeds that have no commercial value continue to be produced despite their potential contribution to crop production, and this situation can be called merciless "seed loss."

Even for major crop species, the quality rate of seeds that are not sorted after harvest is usually less than 90%. The occurrence of defective seeds is partly due to insufficient maturation of the seeds in the mother plant, but damage, deterioration, or contamination during the seed preparation process after harvest can also be a factor. Only after multiple stages of sorting can a high quality rate within the legal standard be achieved, and the large-scale equipment used in such sorting includes spiral density sorters, trommel rotary sorters, vibration sorters, magnetic sorters, weight sorters, and color sorters. However, if the final quality rate after sorting falls below the target, the seeds that have finally been prepared through sowing, cultivation, harvesting, and sorting will just become a pile of garbage. At present, there is no way to improve the quality rate even to the "last 1%" from this stage.

The seed sorting device described above is based on the premise that the quality of a seed is reflected in its outer surface or characteristics detectable from the outside. Characteristics detectable from the outer surface or outside of a seed include shape, surface texture or stickiness, specific gravity, and color. The above-mentioned device is suitable for mass sorting and can prepare most of the harvested seeds to a high-quality state. In contrast, these devices have the disadvantage that it is difficult to fine-tune the operating conditions. Furthermore, when focusing on the biochemical properties of seeds, changes in the embryo and endosperm hidden from the outside can have a much greater impact on the function of the seed than changes in the exposed seed coat. For example, oxidation of endosperm lipids and eating of the embryo by certain pests do not immediately cause changes in the appearance or weight of the seed, but they greatly impair the germination ability. It is presumed that "seed loss" is caused by such inferior seeds.

Near-infrared (NIR) spectroscopy is an effective method for non-destructively evaluating differences and changes in the chemical composition of substances that are primarily composed of organic compounds. In other words, NIR spectroscopy has a high affinity with food, pharmaceuticals, and agricultural products, and its application to their quality control is expanding. Imaging spectroscopy, which has been developed primarily in the field of remote sensing, is also attracting attention as a technology for the quality control of these products, which are characterized by their heterogeneity. Near-infrared imaging spectroscopy, which combines these new technologies, is highly expected to provide the key to eliminating "seed loss."

In fact, there have been reports of the application of NIR imaging spectroscopy to evaluate seed quality in terms of survival, growth potential, genetic purity, and the presence or absence of insect or fungal damage (Non-Patent Documents 1-4). However, the fact that "seed loss" is still occurring clearly indicates that this technology has not yet reached a practical level, or that social implementation has not progressed sufficiently.

Rahman A. et al., Assessment of seed quality using non-destructive measurement techniques: A Review., Seed Science Research, Volume 26, Issue 4, December 2016, pp.285-305. Feng L. et al., Hyperspectral imaging for seed quality and safety inspection: A Review, Plant Methods, 15, Article number: 91 (2019) Venkatesan S. et al., Role of near-infrared spectroscopy in seed quality evaluation: A review. Agric Rev. 2020;41. Long W. et al., A Review of Artificial Intelligence Methods for Seed Quality Inspection based on Spectral Imaging and Analysis, J. Phys.: Conf. Ser., 1769, 012013.

All of the methods reported in the above-mentioned prior art documents enable non-destructive seed selection under certain conditions and are useful. However, all of these prior art techniques are limited to application to specific plant seeds with specific traits, and there remains a strong need to develop new, versatile seed selection methods that can be used for a greater number of varieties and traits.

As a result of intensive research into the above-mentioned problems, the inventors have found that by applying a data set for each plant seed constructed using optical analysis techniques, etc. to multivariate discriminant analysis to generate a large number of candidate discrimination models for the seeds, and when selecting a discrimination model with high predictive accuracy from the candidate models, it is possible to select a model with high predictive accuracy by introducing the concept of "standard selection conditions". They have further researched based on this knowledge, and have thus completed the present invention.
That is, the present invention is as follows.

[1]
A method for selecting a plant seed having a desired trait, comprising the following steps 1 to 7:
(Step 1) Removing a portion of seeds from a seed population;
(a) constructing a data set for the extracted seeds, the data set constructing step including:
(a) irradiating the extracted seeds with light to obtain spectral data and/or derived spectral data;
(b) subjecting the seeds from which the spectral data has been acquired to a quality test for a desired trait and determining a preliminary yield rate of the extracted seeds;
Here, the preliminary yield rate is defined as follows:
[Preliminary quality rate (%)] = [Number of seeds having the desired characteristics among the seeds subjected to quality evaluation] / [Number of seeds subjected to quality evaluation] × 100
(c) Correlating the spectral data of each seed with the results of quality testing;
(Third step) applying multivariate discriminant analysis to the data set constructed in the second step to derive candidates for a good/bad discrimination model for calculating a discriminant score for a desired trait in each seed;
(4th step) The recovery rate and the conformity rate are defined by the following formula:
[Recovery rate (%)] = [Number of seeds selected based on the discriminant score among the seeds subjected to the quality evaluation] / [Number of seeds subjected to the quality evaluation] × 100
[Conformity rate (%)] = [Number of seeds having desired characteristics among the seeds subjected to the quality evaluation and selected based on the discriminant score] / [Number of seeds selected based on the discriminant score among the seeds subjected to the quality evaluation] × 100
In addition, when the selection conditions that make the preliminary non-defective rate and the recovery rate equal are defined as "standard selection conditions",
a step of selecting a discrimination model from the candidates of the quality discrimination model derived in the third step, the step including selecting, as a quality discrimination model, a candidate that satisfies a criterion that the matching rate exceeds a preliminary quality rate under the standard screening conditions;
(Step 5) A step of irradiating the remaining seeds not removed from the seed population in Step 1 and/or other seeds obtained under substantially the same conditions as those of the seeds with light to obtain spectral data, and applying the discrimination model selected in Step 4 to the spectral data to determine a discrimination score for each seed;
(Sixth step) determining a threshold value of the discriminant score based on the discriminant score determined in the fifth step;
(Step 7) A step of recovering seeds predicted to have desired traits by comparing the discrimination scores of each seed determined in step 5 with the discrimination score threshold determined in step 6.
[2]
The selection method according to [1], which is a method for selecting a plant seed having two or more desired traits.
[3]
The selection method according to [1] or [2], wherein the trait is at least one trait selected from the group consisting of germination ability, germination vigor, genotype, stress resistance, dormancy, disease resistance, insect resistance, QTL characteristics, eating quality, heading time, and morphological characteristics.
[4]
The method for selecting a plant according to any one of [1] to [3], wherein the prediction of two or more desired traits is carried out by any one of the following methods (i) to (iii):
(i) A method of scoring seeds individually using a discrimination model for each trait, focusing on the quality of each trait, and predicting that seeds with all scores equal to or above a threshold are good seeds;
(ii) A method of selecting a single discrimination model by determining that a seed is good if all of the desired traits are good, and determining that the other traits are bad, regardless of whether each trait is good or bad, scoring the seeds using the single discrimination model, and predicting that seeds with a score equal to or higher than a threshold are good seeds;
(iii) A method in which a discrimination model for each trait is selected by focusing on the quality of each trait, the scores obtained using the discrimination model for each trait are integrated to obtain an integrated score, and the quality of each seed is predicted based on the ranking of the integrated score.
[5]
The selection method according to any one of [1] to [4], further comprising a step of evaluating the accuracy of the discrimination model.
[6]
The method for selecting a plant according to any one of [1] to [5], wherein the plant is a horticultural crop.

According to the present invention, it is possible to select plant seeds having desired traits from a population of plant seeds having and not having the desired traits non-destructively, highly efficiently, and with high accuracy.

1 is a diagram showing the appearance of vegetable seeds classified by quality category. The numbers below the boxes represent the scores of each seed calculated using the discrimination model in Table 2. 1 is a diagram showing near-infrared reflectance spectra of vegetable seeds classified by quality category, where A to I show the average reflectance spectra for each category, and J shows the reflectance spectrum of an individual seed. FIG. 11 is a flowchart showing a procedure for deriving a pass/fail discrimination model. FIG. 13 is a schematic diagram showing a procedure for deriving a pass/fail discrimination model. This figure shows the relationship between the distribution range of the discrimination score and the discrimination accuracy. A shows a hypothetical situation in which a high-precision discrimination model is used, and B shows a low-precision discrimination model. The preliminary pass rate is set to 80%. Black and gray triangles indicate the preliminary pass rate and the standard threshold value (see Table 3, footnote 2), respectively. Abbreviations: eCDF (empirical cumulative distribution function). (a) Density distribution of the discrimination score for all seeds and eCDF of the proportion of seeds exceeding the horizontal axis score. (b) Density distribution of the discrimination score for good and bad seeds. (c) Distribution range of the discrimination score for good and bad seeds (box plot and scatter plot). (d) Relationship between the setting value of the lower score threshold and discrimination accuracy. (e) Relationship between seed recovery rate and discrimination accuracy. This is a diagram showing the selected wavelength and variable importance in the pass/fail discrimination model. Abbreviations: SPRC (Standardized Partial Regression Coefficient), standard partial regression coefficient; VIP (Variable Importance in Projection), variable importance in projection. This figure shows the relationship between the score lower threshold and the accuracy index in discriminating between good and bad cauliflower seeds. A. Internal validity verification, B. External validity verification. When sequentially predicting germination traits and mating traits using models m-Ca47g and m-Ca47x, respectively, the relationship between the lower threshold setting value for the discrimination scores of both models and the accuracy index in discriminating good seeds (F1 hybrids with normal germination ability) is shown. The gray triangles indicate the standard threshold (see Table 3, footnote 2). Figure 1 shows the internal validity of the good/bad discrimination model. A. 5 pumpkin cultivars, B. 19 pea cultivars, C. 803 lettuce cultivar, D. 22 leek cultivar, E. 94 tomato cultivar, G. 51 leek cultivar, H. 221 tomato cultivar, F, I to K. 47 cauliflower cultivars. F shows the results and accuracy of discrimination for germination traits only, and I shows the results and accuracy of discrimination for hybrid traits only. J shows the results and accuracy of discrimination for good seeds (F1 hybrids with normal germination ability) or not. K shows the results and accuracy of good/bad discrimination by integrated score, reflecting predictions of both traits based on the two discrimination models used in F and I. The symbols, abbreviations, and overview of graphs (a) to (e) are the same as those in Figure 5. Figure 8 shows the external validity of the pass/fail discrimination model. A. Pumpkin variety 5, B. Pea variety 19, C. Lettuce variety 803, D. Leek variety 22, E. Tomato variety 94, F-I. Cauliflower variety 47. A-F correspond to Figures 8A-F, and G-I correspond to Figures 8I-K. Symbols, abbreviations, and the outline of graphs (a)-(e) are the same as those in Figure 5. Figure 1 shows the PR/rPR curves for distinguishing between good and bad cauliflower seeds. A. PR curve, B. rPR curve. When germination traits and mating traits are sequentially distinguished using models m-Ca47g and m-Ca47x, respectively, this shows the relationship between the recall rate in distinguishing each trait and the precision rate (A) and relative precision rate (B) in distinguishing good seeds (F1 hybrids with normal germination ability). Precision rate and relative precision rate are expressed as a curve, and overall discrimination accuracy is expressed as the volume under the curve, but to match the expressions in Figures 11 and 12, these are also referred to as the "curve" and "AUC (area under the curve)". Abbreviations: (r) PR ((relative) Precision-Recall), relationship between (relative) precision rate and recall; AUC (Area Under Curve). 1 is a diagram showing a PR curve for each combination of a pass/fail discrimination model and a data set, Abbreviation: PR (Precision-Recall), the relationship between precision and recall. This figure shows the rPR curve for each combination of the pass/fail discrimination model and the dataset. Relative precision is defined by the following formula: [Relative precision] = (Precision - Initial precision) / (1 - Initial precision). Note that in this specification, "precision" is synonymous with "post-facto precision," and "initial precision" is synonymous with "preliminary precision." The relationship between recall and relative precision is referred to as the "rPR curve," and the area under the curve is referred to as the "rPR-AUC." Figure 13. Pictorial representation of pass/fail discrimination scores: A. Pumpkin variety 5; B. Pea variety 19. Figure 1 shows the germination and early growth characteristics of seeds classified by the quality discrimination score. A and B: pea cultivar 19, C: lettuce cultivar 803, D: leek cultivar 22, E: cauliflower cultivar 47. Image B is an image in which only the pixels corresponding to the green leaves of A have been extracted. The rankings shown in the upper rows of A through D and E are the predicted rankings based on the discrimination model for germination traits, and the rankings shown in the lower row of E are the predicted rankings based on the discrimination model for mating traits.

The present invention will now be described in detail.

The present invention provides a method for selecting a plant seed having a desired trait, the method comprising the following steps 1 to 7 (hereinafter, sometimes referred to as the "method of the present invention"):
(Step 1) Removing a portion of seeds from a seed population;
(a) constructing a data set for the extracted seeds, the data set constructing step including:
(a) irradiating the extracted seeds with light to obtain spectral data and/or derived spectral data;
(b) subjecting the seeds from which the spectral data has been acquired to a quality test for a desired trait and determining a preliminary yield rate of the extracted seeds;
Here, the preliminary yield rate is defined as follows:
[Preliminary quality rate (%)] = [Number of seeds having the desired characteristics among the seeds subjected to quality evaluation] / [Number of seeds subjected to quality evaluation] × 100
(c) Correlating the spectral data of each seed with the results of quality testing;
(Step 3) applying multivariate discriminant analysis to the data set constructed in Step 2 to derive candidates for a good/bad discrimination model for calculating a discriminant score for a desired trait in each seed;
(4th step) The recovery rate and the conformity rate are defined by the following formula:
[Recovery rate (%)] = [Number of seeds selected based on the discriminant score among the seeds subjected to the quality evaluation] / [Number of seeds subjected to the quality evaluation] × 100
[Conformity rate (%)] = [Number of seeds having desired characteristics among the seeds subjected to the quality evaluation and selected based on the discriminant score] / [Number of seeds selected based on the discriminant score among the seeds subjected to the quality evaluation] × 100
In addition, when the selection conditions that make the preliminary non-defective rate and the recovery rate equal are defined as "standard selection conditions",
a step of selecting a discrimination model from the candidates of the quality discrimination model derived in the third step, the step including selecting, as a quality discrimination model, a candidate that satisfies a criterion that the matching rate exceeds a preliminary quality rate under the standard screening conditions;
(Step 5) A step of irradiating the remaining seeds not removed from the seed population in Step 1 and/or other seeds obtained under substantially the same conditions as those of the seeds with light to obtain spectral data, and applying the discrimination model selected in Step 4 to the spectral data to determine a discrimination score for each seed;
(Sixth step) determining a threshold value of the discriminant score based on the discriminant score determined in the fifth step;
(Step 7) A step of recovering seeds predicted to have desired traits by comparing the discrimination scores of each seed determined in step 5 with the discrimination score threshold determined in step 6.

In the method of the present invention, the term "plant seed" refers to the seeds of any plant, and is not particularly limited. The plant is not particularly limited as long as it is a seed plant, and may be, for example, either angiosperms or gymnosperms. Furthermore, when the plant is angiosperms, the plant may be either dicotyledonous or monocotyledonous. Furthermore, when the plant is dicotyledonous, the plant may be either sympetalous or polypetalous. In one aspect, the method of the present invention can be used to select seeds of plants with high added value. Examples of such plants with high added value include, but are not limited to, horticultural crops and plants that can be used as building timber. Horticultural crops include vegetables, fruit trees, and ornamental plants. Examples of vegetables include, but are not limited to, pumpkin, peas, lettuce, green onions, tomatoes, cauliflower, bitter melon, okra, onions, Japanese ginger, soybeans, butterbur, asparagus, Chinese chives, broad beans, celery, carrots, mizuna, komatsuna, chrysanthemum, radish, broccoli, spinach, Chinese cabbage, arugula, lotus root, turnip, avocado, cucumber, paprika, garlic, corn, zucchini, parsley, cilantro, eggplant, and green peppers. Examples of fruit trees include, but are not limited to, plum, fig, akebia, acerola, avocado, olive, orange, persimmon, quince, guava, cranberry, walnut, grapefruit, cherry, and pomegranate. In addition, examples of flowers include, but are not limited to, morning glory, cockscomb, cosmos, zinnia, columbine, globe amaranth, petunia, periwinkle, cabbage, sunflower, impatiens, portulaca, portulaca, balloon vine, marigold, gypsophila, snapdragon, calendula, sweet pea, stock, dianthus, daisy, nigella, nemesia, nemophila, poppy, verbena, pansy, viola, corn poppy, cornflower, lupine, and forget-me-not. In addition, examples of plants that can be used as building timber include, but are not limited to, cedar, cypress, Japanese cypress, chestnut, zelkova, cherry, beech, walnut, falcata, and red pine.

In the method of the present invention, the term "desired trait" encompasses both traits that the seed itself possesses and traits that are expressed in a plant that germinates from the seed when the seed is grown. Examples of "traits" that are the subject of selection in the method of the present invention include, but are not limited to, germination ability, germination vigor, genotype, stress resistance, dormancy, disease resistance, insect resistance, QTL characteristics, taste, heading time, and morphological characteristics such as leaf size.

(First step)
The first step of the present invention is characterized in that some seeds are taken out from the seed population. The number of seeds constituting the seed population may be 2 or more, with no particular upper limit. For example, the seed population may be, but is not limited to, a seed population consisting of usually 2 to 10,000,000 seeds, preferably 1,000 to 10,000,000 seeds, and more preferably 10,000 to 10,000,000 seeds. Alternatively, the seed population may be expressed by weight. For example, the seed population may be, but is not limited to, a seed population weighing 1 g to 10,000 kg, preferably 1 kg to 10,000 kg, and more preferably 10 kg to 10,000 kg.

　The "portion" of the seed population may vary depending on the "traits" to be selected, but is usually, but not limited to, about 50 to 50,000 seeds, preferably 100 to 30,000 seeds, and more preferably 200 to 3,000 seeds.

(Second step)
The second step of the present invention is to construct a data set for the extracted seeds. The construction of the data set includes at least the following steps (a) to (c):

(a) irradiating the extracted seeds with light to obtain spectral data and/or derived spectral data.
(b) subjecting the seeds for which the spectral data has been acquired to a quality inspection for a desired characteristic, and determining a preliminary yield rate of the extracted seeds. The preliminary yield rate in (b) is defined by the following formula:
[Preliminary quality rate (%)] = [Number of seeds having the desired characteristics among the seeds subjected to quality evaluation] / [Number of seeds subjected to quality evaluation] × 100
(c) Correlation of the spectral data of each seed with the results of quality testing.

In this specification, the "preliminary conforming rate" may be referred to as the "initial conforming rate."

In (a) of constructing the data set, the method of irradiating light onto the seeds to obtain spectral data (reflection, absorption, and/or transmission spectral data) and/or derived spectral data thereof may be a method generally used in the technical field of optical analysis. The light irradiated onto the seeds is not particularly limited as long as it can obtain spectral data. The light irradiated onto the seeds may be, for example, microwaves, terahertz waves, infrared light, visible light, ultraviolet light, X-rays, and gamma rays, but is not limited to these. In a preferred embodiment of the present invention, the light may be visible light or infrared light. The method of obtaining the spectral data may be, for example, the method and conditions used in the examples of this application, as well as the method taught in Patent No. 6782408, but is not limited to these.

Furthermore, the method of generating derived spectral data from the obtained spectral data may be a method known per se in the technical field of optical analysis. For example, the derived spectral data can be generated by performing reciprocal (1/R) transformation, logarithmic transformation, standard normalization, smoothing and smoothing differentiation using an SG (Savizky-Golay) filter, or any combination of these on the obtained spectral data (R). Note that only one of the spectral data and the derived spectral data may be used, or two or more of them may be used. In a preferred embodiment, the spectral data used may be one of the spectral data and the derived spectral data generated therefrom that is capable of deriving the most accurate prediction model.

In addition, the equipment used to acquire the spectral data may be any known equipment. For example, the spectral characteristics of biological tissues can be measured by point measurement using a fiber optic spectrometer, or by using a hyperspectral camera, which is a type of remote sensor, to measure (measure) them together with coordinate information (images) (surface measurement). When making measurements, an equipment that exhibits high detection sensitivity in the target wavelength range may be appropriately selected. In the method of the present invention, for example, equipment equipped with a photodetector such as a CCD, CMOS, CQD, InGaAs, HgCdTe (MCT), or Type II superlattice (T2SL) may be used, but is not limited to these. In addition, when acquiring the spectral data, a reflectance correction image may be generated and a seed recognition model may be applied, if necessary (see ST1-3 in FIG. 3).

The quality inspection in (b) of constructing the data set is also not particularly limited as long as the preliminary pass rate for the desired trait can be determined. For example, if the desired trait is "germination ability," this can be easily confirmed by subjecting the seeds from which the spectral data has been acquired to a germination test that complies with the International Seed Inspection Standards established by ISTA or a similar germination test. If the number of seeds subjected to the germination test is A (pieces) and the number of seeds that actually germinated is B (pieces), the [preliminary pass rate (%) (i.e., germination rate)] can be determined as 100B/A (%) from the definition formula [number of germinated seeds (B pieces) among the seeds subjected to quality evaluation] / [number of seeds subjected to quality evaluation (A pieces)] × 100.

　To explain "matching the spectral data of each seed with the quality test results" in (c) of constructing the dataset specifically using the case where the trait is "germination ability", the spectral data (reflection, absorption, or transmission spectral data or derivative spectral data) of each seed obtained by irradiating light on seeds A, B, and C are SA, SB, and SC, respectively. Also, when seeds A, B, and C are subjected to a quality test and the test results show that seed A germinates normally, seed B germinates abnormally, and seed C does not germinate, in (c), the spectral data is intended to be matched with the quality results, such as SA being normal germination, SB being abnormal germination, and SC being non-germination. At this time, dummy variables may be assigned in order to make it easier to determine whether the result is good or bad. For example, in the case of "germination ability", variables such as 1 (good) can be assigned to normal germination and -1 (bad) can be assigned to abnormal germination and non-germination, but this is not limiting.

(Third process)
The third step of the present invention is a step of deriving candidates for a good/bad discrimination model for calculating a discrimination score for a desired trait in each seed by applying multivariate discriminant analysis to the data set constructed in the second step.

The multivariate discriminant analysis in this step may be performed using a method known per se, and any method may be used as long as the desired effect of the present invention can be obtained. In the method of the present invention, the multivariate statistical analysis technique may be, but is not limited to, partial least squares (PLS) or principal component analysis (PCA). The multivariate discriminant analysis may preferably be partial least squares-Discriminant analysis (PLS-DA).

As an example, we will outline how to perform multivariate discriminant analysis using PLS-DA.

PLS-DA involves modeling an equation such as the following Equation 1 using the explanatory variables and target variables that have been set.

(Equation 1)
y=a+b ₁ x ₁ +b ₂ x ₂ +b ₃ x ₃ +b ₄ x ₄ +...

Here, x ₁ , x ₂ , x ₃ and x ₄ are explanatory variables, y is a response variable, a is an intercept (constant), and b ₁ , b ₂ , b ₃ and b ₄ are partial regression coefficients (constants).

The spectral data acquired in the second step is used as explanatory variables, and the results obtained by the quality inspection are used as objective variables, and a candidate quality discrimination model for calculating a discrimination score for the desired traits in each seed is derived. The method for deriving the candidate quality discrimination model can also be a method known per se, but the method used in the examples of the present invention is described below as an example.

From the reflection, absorption, or transmission spectral data (R) obtained by irradiating light on the seeds, including the data itself (X1), derived spectral data Xs is generated using the above-mentioned conversion method and any combination thereof. In the discriminant analysis, an attempt is made to use all of the derived spectra as explanatory variables, but if the predictive performance of the derived discriminant model is equivalent, the simplest derived pattern can be adopted.

(Procedure for deriving candidate pass/fail discrimination models)
Final preparation of the dataset is as follows:
A reflection, absorption, or transmission spectrum is acquired, and a derived spectrum is generated to prepare explanatory variable data. If necessary, standardization is performed for either or both of the objective variable and the explanatory variable. A discriminant model is derived by multivariate discriminant analysis. For example, a discriminant model can be derived by linear sparse modeling as follows. For all combinations of objective variable and explanatory variable data, first, an initial value of the weight of each explanatory variable is determined by PLS regression or Ridge regression. Next, a solution path that represents the relationship between the regularization coefficient λ and the partial regression coefficient, and the number of explanatory variables whose partial regression coefficients are not 0, is calculated by Adaptive LASSO regression. If the number of explanatory variables that minimize the residual sum of squares in the regression is p, candidates for combinations of explanatory variables suitable for reducing the number of explanatory variables to p or less can be determined from the solution path. The final discriminant model is derived by PLS-DA. A discriminant model is derived using 2 to p explanatory variables determined from the solution path in Adaptive LASSO regression. The derived discriminant model is set as a candidate for the good/bad discrimination model. Note that at this stage, many candidate discriminant models are presented. The discriminant model that gives a higher relative score to good seeds is selected in the next step 4.

(Fourth step)
The fourth step of the present invention is a step of selecting a discrimination model from the candidates of the pass/fail discrimination model derived in the third step.

In the fourth step, the "recovery rate," "conformance rate," and "standard sorting conditions" are defined as follows:

[Recovery rate (%)] = [Number of seeds selected based on the discrimination score from among the seeds subjected to quality evaluation] / [Number of seeds subjected to quality evaluation] x 100

[Conformity rate (%)] = [Number of seeds with the desired traits among the seeds subjected to quality evaluation and selected based on the discrimination score] / [Number of seeds selected based on the discrimination score among the seeds subjected to quality evaluation] x 100

[Standard sorting conditions]: Sorting conditions that make the [preliminary quality rate] and [recovery rate] equal.

When the recovery rate, precision rate, and standard selection conditions are defined as described above, candidates that meet the criterion that the precision rate exceeds the preliminary pass rate are selected as pass/fail discrimination models, making it possible to select a discrimination model capable of discriminating pass/fail for the desired trait from the many model candidates obtained in the third step. Note that the greater the precision rate exceeds the preliminary pass rate, the better the discrimination model is considered to be. Furthermore, in order to enable easy adjustment of the precision rate as needed, it is preferable to select a discrimination model that has a high rank correlation between increases and decreases in the recovery rate and increases and decreases in the precision rate (i.e., a relationship in which a decrease in the recovery rate reliably increases the precision rate, and an increase in the recovery rate reliably decreases the precision rate).

In this specification, "conformance rate" may be referred to as "post-test quality rate."

(Fifth step)
The fifth step of the present invention is a step of irradiating the remaining seeds not extracted from the seed population in the first step and/or other seeds obtained under substantially the same conditions as the seeds to obtain spectral data, and determining the discrimination score of each seed by applying the discrimination model selected in the fourth step to the spectral data. In other words, the fifth step is a step of obtaining spectral data for non-training seeds in the seed population, and determining the discrimination score of each seed by applying the discrimination model obtained in the fourth step to the spectral data. The method and conditions for obtaining the spectral data may be the same as those used for the training seeds, and may be partially different as long as the desired effect is obtained.

(Sixth step)
The sixth step of the present invention is a step of determining a threshold value of the discrimination score based on the discrimination score determined in the fifth step. If the pre-quality rate in the seed population is "k%," the score that is "k%" from the top can be set as a threshold value (standard threshold value) for standard selection. In addition, if a discrimination model having a high rank correlation between the increase/decrease in the recovery rate and the increase/decrease in the post-quality rate is selected in the fourth step described above, the matching rate can be adjusted by appropriately increasing/decreasing the threshold value based on the standard threshold value.

(Seventh step)
The seventh step of the present invention is a step of recovering seeds predicted to have the desired trait by comparing the discrimination scores of each seed determined in the fifth step with the discrimination score threshold determined in the sixth step. The recovery of the seeds may be achieved by selecting and removing the seeds predicted to have the desired trait, or by selecting and removing the seeds predicted not to have the desired trait, and recovering the seeds not removed as seeds having the desired trait.

In one embodiment of the present invention, the method of the present invention can be a method for selecting plant seeds having two or more desired traits.

For example, examples of two or more traits include, but are not limited to, a combination of two or more of the traits listed above (i.e., germination ability, germination vigor, genotype, stress resistance, dormancy, disease resistance, insect resistance, QTL characteristics, eating quality, heading time, and morphological characteristics such as leaf size, etc.).

When the method of the present invention is a method for selecting two or more desired traits, the prediction of the two or more desired traits is performed by any one of the following methods (i) to (iii).

(i) A method that focuses on the quality of each trait, scores seeds individually using a discrimination model for each trait, and predicts that seeds with all scores above a threshold are good ("sequential prediction using multiple models").

(ii) A method in which a single discriminant model is selected by ignoring the quality of each trait, determining that seeds with all desired traits in good condition are good, and determining all other traits as bad, scoring the seeds using that single discriminant model, and predicting that seeds with scores above a threshold are good seeds ("direct prediction using a single model").

(iii) A method in which a discrimination model for each trait is selected by focusing on the quality of each trait, the scores obtained using the discrimination model for each trait are integrated to obtain an integrated score, and the quality of each seed is predicted based on the ranking of the integrated score ("direct prediction using an integrated model").

Using either method, selection of seeds with two or more desired traits can be achieved.

In one aspect of the present invention, the method of the present invention may further include a step of evaluating the accuracy of the discrimination model. The accuracy evaluation is an evaluation of the "generalized performance" of the selected discrimination model. Specifically, the accuracy evaluation of the pass/fail discrimination model can be performed by the following method.

The discrimination model obtained by carrying out the above-mentioned steps will undergo internal validation when applied to the internal data used in its derivation (i.e., the seed population from which the dataset was constructed) and external validation when applied to external data obtained independently of the derivation of the model.

To evaluate the overall accuracy of a pass/fail discrimination model that is not dependent on the set value of the lower score threshold, ROC (Receiver Operating Characteristic) curves and PR (Precision-Recall) curves can be used. In particular, in quality inspections where misclassification, in which the adoption of defective products (false positives) is considered more serious than the rejection of good products (false negatives), PR curves, which are sensitive to false positives, are often used.

In both ROC and PR curves, a discrimination model is considered to be superior when its AUC ( area under the curve ) is closer to the maximum value of 1. However, while the minimum value of ROC-AUC is always 0.5, the minimum value of PR-AUC is the initial precision rate, i.e., the prior quality rate, and this varies depending on the data set to which it is applied.

PR-AUC is suitable for verifying which of a number of discriminant models shows superior performance for a specific data set, but is not suitable for verifying whether a single discriminant model shows equivalent performance for different data sets. This is because it is affected by the prior pass rate.

To compensate for the above-mentioned shortcomings of the PR curve, we define the relative precision, which is expressed by the following formula:

[Relative precision] = (precision - initial precision) / (1 - initial precision)

The relationship between relative precision and recall is called the rPR (relative Precision-Recall) curve, and the area under the curve is called the rPR-AUC. rPR-AUC does not depend on the dataset to which it is applied, and always takes a value in the closed interval [0, 1].

When n models are used in pass/fail discrimination, the relationship between precision and relative precision and recall forms a hypersurface occupying an n+1-dimensional space. For convenience, the relationships between precision and relative precision and recall are referred to as PR curves and rPR curves, and the ratios below the hypersurface they form are referred to as PR-AUC and rPR-AUC, regardless of the number of discrimination models used.

For all derived discrimination models, PR and rPR curves are created when each is applied to the data set derived from the target crop species. In addition, PR-AUC and rPR-AUC are calculated from each curve. The accuracy of the discrimination model is evaluated from the shape of the created curve and the AUC value.

The present invention will be explained in more detail in the following examples, but the present invention is not limited to these examples.

[Identifying vegetable seeds with quality that meets shipping requirements]
(Materials) Vegetable seeds (pumpkin, pea, lettuce, green onion, tomato, and cauliflower) harvested from 2015 to 2022 provided by Tokita Seed Co., Ltd. were used as test materials (Table 1).

The seeds were irradiated with near-infrared light by applying a DC voltage to two 24V 250W halogen lamps with aluminum mirrors (JTR24V250W10H/5-AL, GX5.3 base, manufactured by Kahoku Lighting Solutions), and near-infrared hyperspectral images were taken using a line-scan hyperspectral camera (CV-N801HS, manufactured by Sumitomo Electric Industries, Ltd.). The near-infrared lens was an image-side telecentric lens (manufactured by Sumitomo Electric Industries, Ltd.) with a focal length of 30 mm. The working distance during imaging was 28 cm, the spatial resolution was 90 ppi, and the wavelength sampling interval was 6 nm. The wavelength sensitivity range of the camera used was 980 to 2,350 nm. The reflectance at each wavelength was calibrated to the equivalent of diffuse reflectance based on the images of a standard reflector with a reflectance of 99% and dark current (reflectance of 0%) that constitute a contrast target (SRT-MS-050, Labsphere, USA). The reflectance spectrum of the seeds was the average of the reflectance spectra recorded for each pixel in the area occupied by each seed.

The appearance of the seeds was photographed using an 8k color line scan camera (e2v EV71C4CCL8005-BA0) under illumination by a white LED light source (Leimac IDBA-HMS150WHV-S). The lens was an object-side telecentric lens (Optoart FT04-150CL) with an optical magnification of 0.4x, and the spatial resolution was 2,032 ppi. A standard reflector with a reflectance of 25% (when the seed surface was dark) or 50% (when the seed surface was light) that constitutes the contrast target (SRT-MS-050 mentioned above) was used to adjust the white balance of the camera.

(method)
I. Construction of Data Set (1) Seeds were uniformly irradiated with light from a halogen lamp, and near-infrared hyperspectral images containing spectral information in the near-infrared wavelength range were taken. For some seeds, high-resolution external images were also taken with an 8k line camera (Figure 1).

(2) The near-infrared reflectance spectrum of the area occupied by each seed was calculated from the near-infrared hyperspectral image (Figure 2).

(3) The seeds were subjected to germination tests in accordance with the international seed testing regulations established by the International Seed Testing Association (ISTA), and their germination characteristics were classified into three groups: normal germination, abnormal germination, and no germination.

(4) Cauliflower variety 47 (variety number omitted below) is used as an F1 hybrid, but due to incomplete self-incompatibility, self-fertilized seeds may be formed. The hybridization characteristics of each seed, whether F1 or self-fertilized, were determined by isozyme type analysis using seedlings.

(5) Dummy variables were assigned to both good and bad seeds, and these were used as the objective variables in the discriminant analysis described below. When discriminating between germination characteristics, the value of the dummy variable was set to 1 (good) for normally germinated seeds and -1 (bad) for others. When discriminating between cauliflower hybridization characteristics, F1 hybrids were assigned a value of 1 and others -1. When discriminating between germination and hybridization characteristics that meet the requirements for good seeds, F1 hybrids with normal germination ability were assigned a value of 1 and others -1.

II. Derivation of a model for discriminating between good and bad vegetable seeds (6) A discriminant analysis was performed to predict whether a seed has the desired traits (normal germination for all crop species, and F1 hybrid for cauliflower) from the near-infrared reflectance spectrum of the seed.

(7) The procedure for the discriminant analysis was based on the method described in JP 2023-125301 A. In the present invention, smoothed differentiation of the near-infrared reflectance spectrum and column combination of spectra with different differential orders were omitted (Figure 3).

(8) The near-infrared reflectance spectrum (R), which serves as the explanatory variable, was subjected to reciprocal (1/R) and logarithmic (pseudo absorbance, -logR) transformation, as well as standard normalization (SNV, Standard Normal Variate; effective wavelength range 980-2,200 nm) and smoothing using a Savizky-Golay (SG) filter (five-point, third-order polynomial approximation) in arbitrary combinations to generate 11 derived spectra. A total of 12 spectra, including the original spectrum, were used as explanatory variables (Figure 4).

(9) A model for discriminating between good and bad seeds was derived using a multivariate linear sparse modeling technique that combines Adaptive LASSO (Least Absolute Shrinkage and Selection Operator) and PLS-DA (Partial Least Squares-Discriminant Analysis). At this stage, many candidate models are presented, but it is necessary to select the one that gives a higher relative score to good seeds. The model selection method is specifically shown below.

III. Selection of a Good/Bad Discrimination Model (10) In order to devise a method for selecting a desirable discrimination model from among the many candidates presented above, the following thought experiment was carried out.

1) Assume a high-precision discrimination model that gives clearly separated scores to good and bad seeds, and a low-precision discrimination model that gives widely overlapping scores. Here, the pre-qualifying rate is set to 80%, and the former gives scores to good and bad seeds that follow a normal distribution with means of 1 and -1 and standard deviations of 0.25, while the latter gives scores to good and bad seeds that follow a normal distribution with means of 0.05 and -0.05 and standard deviations of 0.25 (Figure 5).

2) A sorting task is performed in which seeds are collected in descending order of score. When a high-precision discrimination model is used, the precision rate (posterior quality rate) remains close to 100% until the collection rate reaches 80%, which is the same as the prior quality rate (Figure 5A(e)). On the other hand, when a low-precision discrimination model is used, the precision rate quickly converges to the prior quality rate (Figure 5B(e)). In either case, when the collection rate matches the prior quality rate, the precision rate and recall rate (recovery rate of good seeds) become equal.

3) When selecting good seeds using a discrimination model, after all the seeds have been measured, the seeds with the highest scores are not collected, but in practice, seeds with scores above a predetermined lower threshold are collected one by one. Since the relationship between the score and its ranking can be easily calculated, the relationship between each index related to discrimination accuracy, including the precision rate and recall rate, and the recovery rate (Figure 5(e)) can be redrawn as the relationship with the discrimination score (Figure 5(d)).

4) Here, the sorting conditions that make the preliminary non-defective rate and the recovery rate equal are defined as the "standard sorting conditions," and the lower threshold of the discrimination score of the seeds to be recovered at that time is defined as the "standard threshold."

From the above, it can be seen that a requirement for an excellent discrimination model is that when selection is performed using standard selection conditions, the precision rate (i.e., the ex-post rate of acceptable products) and its equivalent, the recall rate, are significantly higher than the ex-ante rate of acceptable products.

(11) In addition, when selecting good seeds, it is desirable to be able to adjust the conformance rate as necessary. In other words, it is more desirable to have a relationship in which the conformance rate can be increased with high accuracy as the recovery rate is lowered (the lower threshold score is increased).

(12) The quality discrimination models for each crop seed shown in Tables 2-1, 2-2, and 2-3 below were selected based on the requirement that they satisfy the above two points. Figure 6 shows the relationship between the wavelength bands adopted as necessary explanatory variables in each model in the same table and the standard partial regression coefficients and variable importance in projection (VIP).

IV. Combined Processing of Multiple Discriminant Scores Embodiments in which multiple traits are assessed are described hereafter.

(13) When good cauliflower seeds are defined as those that (1) have normal germination ability and (2) are F1 hybrids, the following two methods can be used to distinguish good cauliflower seeds.

Method 1) The quality of germination and mating characteristics are scored individually using two models (m-Ca47g and m-Ca47x), and seeds with the highest scores are predicted to be good seeds. Method 2) The quality of individual characteristics is ignored, and only good seeds are predicted (m-Ca47).

(14) In method 1), to obtain the desired discrimination accuracy, it is necessary to set a lower score threshold for the two models. When n discrimination models are used, the relationship between the setting value of the lower score threshold and each index related to discrimination accuracy forms the shape of an n+1-dimensional hypersurface. When n=2 as in this example, each index forms a curved surface, but it can also be drawn two-dimensionally as a heat map (Figure 7). However, this method is significantly less easy to operate than method 2).

(15) In order to integrate the scores of multiple discriminant models and set only one lower threshold, we devised the following mathematical process.

1) Let M1 through Mk be models that discriminate between good and bad traits 1 through k, and y1(X) through yk(X) be the discriminant scores obtained by substituting explanatory variable X into these models.

2) A sufficient number n (several hundred) of seeds that make up a particular lot are measured, and scores y1(x) to yk(x) are obtained from the explanatory variables x (x1 to xn) that correspond to each. If the pre-qualification rates for each trait are assumed to be p1 to pk, then the n×p (p1 to pk)-th score from the top of y (y1 to yk) is defined as yp (yp1 to ypk). In addition, the standard deviation of y (y1 to yk) is defined as s (s1 to sk).

3) Score y (y1 to yk) is converted to score z (z1 to zk) using the following formula:
z=(y-yp)/s

4) The minimum value of score z (z1 to zk) is the combined score for discrimination using M1 to Mk.

5) The quality of each seed is judged based on the ranking of the integrated score. Standard conditions and standard thresholds can also be determined when using the integrated score for judgment.

This process was applied to m-Ca47g and m-Ca47x to generate model m-Ca47gx (Table 3-1 and Table 3-2).

V. Accuracy evaluation of the good/bad discrimination model (16) For all selected discrimination models, the distribution area of the discrimination scores for all seeds or good/bad seeds, and the relationship between the set value of the score lower threshold or the seed recovery rate and each index related to the discrimination accuracy are shown in Figures 8 and 9. The configuration of the figures is the same as that of Figure 5.

(17) Figures 8 and 9 respectively correspond to the results of internal validity verification when each discriminant model is applied to the internal data used in its derivation, and external validity verification when it is applied to external data obtained independently of the model derivation. Figure 9 shows the evaluation results of the generalization performance of the model.

(18) ROC (Receiver Operating Characteristic) curves and PR (Precision-Recall) curves are often used to evaluate the overall accuracy of a pass/fail discrimination model that is independent of the set value of the lower score threshold. In particular, in quality inspections in which the adoption of defective products (false positives) is considered more serious than the rejection of good products (false negatives), the PR curve, which is sensitive to false positives, is often used.

(19) In both the ROC curve and the PR curve, a discrimination model with an AUC ( area under the curve) value closer to 1, which is the maximum value, is considered to be superior. However, while the minimum value of the ROC-AUC is always 0.5, the minimum value of the PR-AUC is the initial precision rate, i.e., the prior quality rate, and varies depending on the data set to which it is applied.

(20) PR-AUC is suitable for verifying which of a number of discriminant models shows superior performance for a specific data set, but is not suitable for verifying whether a single discriminant model shows equivalent performance for different data sets. This is because it is affected by the prior pass rate.

(21) To compensate for the above-mentioned shortcomings of the PR curve, we define the relative precision, which is expressed by the following formula:

[Relative precision] = (precision - initial precision) / (1 - initial precision)

(22) The relationship between relative precision and recall is called the rPR (relative Precision-Recall) curve, and the area under the curve is called the rPR-AUC. The rPR-AUC does not depend on the dataset to which it is applied, and always takes a value in the closed interval [0, 1].

(23) When n models are used in pass/fail discrimination, the relationship between the precision rate or relative precision rate and recall rate also forms an n+1-dimensional hypersurface shape. For convenience, the relationships between the precision rate and relative precision rate and recall rate are referred to as the PR curve and rPR curve, and the ratios below the hypersurface they form are referred to as the PR-AUC and rPR-AUC, regardless of the number of discrimination models used.

(24) For all derived discrimination models, the PR curves (Figures 10A and 11) and rPR curves (Figures 10B and 12) are shown when applied to the data sets derived from the target crop species. In addition, the PR-AUC and rPR-AUC calculated from each curve are shown in Tables 3-1 and 3-2.

VI. Selection of good seeds based on a discrimination model (25) For near-infrared hyperspectral images of seeds, we created software that calculates the scores according to the discrimination model shown in Table 2 for each pixel or area occupied by the seed, converts them into pseudocolors or brightness values, and visualizes them (FIG. 13).

(26) Using the above software, seeds with high and low discrimination scores related to germination or mating characteristics were collected. For each category, seed germination and early growth were evaluated (Figure 14), and the isozyme type of cauliflower was analyzed to determine the ex-post quality rate (suitability rate) (Table 4).

(Results and Discussion)
(1) For all crop species, it was impossible to subjectively determine the quality of the seeds after the germination test from their pre-germination appearance and near-infrared reflectance spectrum (Figures 1 and 2).

(2) When the derived discrimination model was applied to the target varieties, it was confirmed that the distribution of scores, without exception, was in the order from high to low, from good seeds to bad seeds (Figure 8 and Figures 9(b) and (c)).

(3) In the model predicting germination traits, the distribution range of scores shown by abnormally germinated seeds was located between the scores shown by normally germinated and non-germinated seeds (Figures 8A-H and Figures 9A-F (b) and (c)), and it was inferred that the discrimination scores generally reflect the growth potential of the seeds.

(4) The distribution range of the discrimination scores overlapped between good and bad seeds, and misclassification could not be avoided for any of the crop species. In addition, compatibility of the discrimination model was confirmed between tomato varieties 94 and 221, but it was suggested that the discrimination accuracy was inferior compared to when applied to the same variety (Table 3 and Figures 12H and I). In the case of leeks, it was determined that sharing the discrimination model between varieties 22 and 51 was not practical (Table 3 and Figures 12F and G).

(5) The distribution range of the discrimination scores was not the same on different test occasions, even when the same variety and lot of seeds were used (Fig. 8, Fig. 9 (a) to (c)); different lots of tomato variety 94 were used in Fig. 8 and Fig. 9). The derived discrimination model was capable of repeated application, at least to the same variety, but it was desirable to optimize the lower threshold score for selecting good seeds for each occasion.

(6) When the discrimination model is applied to the near-infrared spectra obtained from randomly selected seeds and the scores are calculated, the quality of each seed is unknown. In addition, the density distribution of the scores shows only a featureless single peak (Figure 9C, E-I(a); A, B, and D are not randomly selected), and it is difficult to estimate the lower threshold score for desired sorting from this shape. On the other hand, it is possible to determine a standard threshold for recovering seeds with the same rate as the estimated prior quality rate, and it is reasonable to define this as the standard sorting condition (standard condition). Therefore, it is desirable to understand the prior quality rate of the seeds to be used as accurately as possible.

(7) If you want to increase the ex-post quality rate (matching rate) more than under the standard conditions, you can increase the lower threshold of the discrimination score. If you want to increase the recall rate (recovery rate of good seeds), you can lower the lower threshold of the score.

(8) In the above, it is not easy to estimate how much changing the lower threshold of the discriminant score will affect the precision rate and recall rate. The transformation of subtracting a standard threshold from the score and then dividing by the standard deviation of the score has the advantage of making the variance in the scores uniform, just like standardizing (normalizing) variables. This is also effective when selecting good seeds using a single model.

(9) The score conversion described above makes it possible to consolidate the thresholds to be adjusted into one even when selecting good seeds using multiple discrimination models.

(10) In cauliflower, when predicting F1 hybrids with normal germination ability, which are good seeds, higher discrimination accuracy was observed when a combination of models predicting germination and mating characteristics individually was used rather than using a single model that directly discriminated between good and bad seeds (Table 3 and Figures 12L and M).

(11) When a model predicting germination traits was applied and the germination and early growth of seeds with high and low discrimination scores were compared, it was found that the seeds with high scores not only had a high germination rate but also tended to show vigorous growth after germination (Figure 14). The same tendency was observed in lettuce, for which the derived quality discrimination model performed somewhat poorly. From this, it was inferred that the discrimination scores for germination traits reflect the growth potential of the seeds.

(12) By confirming that the precision rate (synonymous with the ex-post precision rate and identical to the recall rate) under standard selection conditions was sufficiently higher than the ex-ante precision rate, it was possible to efficiently narrow down the number of candidate discrimination models to a preferred model (Table 3 and Figure 3). Furthermore, by defining the relative precision rate and calculating the area under the rPR curve (rPR-AUC), which represents the relationship between this and the recall rate, it became possible to objectively compare the discrimination accuracy regardless of the combination of model and dataset (Table 3 and Figure 12).

(13) With the technology of the present invention, it is possible to determine whether the seeds are good or bad within about one day of receiving the results of the seed quality evaluation test, and if possible, a specific discrimination model can be obtained. Therefore, even if the occurrence of defective seeds is accidental and not caused by the genetic properties of the variety, etc., it is possible to start the selection process to restore the good seed rate to a level sufficient for shipping standards shortly after the problem is discovered.

(14) To make the above possible, it is necessary to obtain near-infrared hyperspectral images of the seeds in advance that can be linked to the results of quality evaluation tests.

(15) As mentioned above, deriving a discrimination model does not require a great deal of time or effort. It is possible to continue updating a model applied to a specific crop variety to improve its accuracy, with the intention of reusing it on another occasion; however, the reasons for the occurrence of defective seeds often differ from occasion to occasion. It is important not only to enrich the library of discrimination models compatible with various crop varieties, but also to establish a seed quality control system that can derive the optimal good/bad discrimination model for each occasion when a problem occurs and respond immediately to resolve it.

(16) Near-infrared hyperspectral images can be acquired all at once, and the data obtained from them can be used to distinguish multiple traits, such as seed germination characteristics and genotype. The images contain information about grain size and are also effective in estimating grain weight. The present invention provides a means of performing advanced seed sorting at low cost and in a small space, without introducing and installing multiple sorting machines with different functions.

According to the present invention, it is possible to select plant seeds having desired traits from a population of plant seeds having and not having the desired traits non-destructively, highly efficiently, and with high accuracy. Therefore, the present invention is extremely useful, for example, in the seed and seedling industry.

This application is based on patent application No. 2023-089036 filed in Japan (filing date: May 30, 2023), the contents of which are incorporated in their entirety into this specification.

Claims (6)

A method for selecting a plant seed having a desired trait, comprising the following steps 1 to 7:
(Step 1) Removing a portion of seeds from a seed population;
(a) constructing a data set for the extracted seeds, the data set constructing step including:
(a) irradiating the extracted seeds with light to obtain spectral data and/or derived spectral data;
(b) subjecting the seeds from which the spectral data has been acquired to a quality test for a desired trait and determining a preliminary yield rate of the extracted seeds;
Here, the preliminary yield rate is defined as follows:
[Preliminary quality rate (%)] = [Number of seeds having the desired characteristics among the seeds subjected to quality evaluation] / [Number of seeds subjected to quality evaluation] × 100
(c) Correlating the spectral data of each seed with the results of quality testing;
(Step 3) applying multivariate discriminant analysis to the data set constructed in Step 2 to derive candidates for a good/bad discrimination model for calculating a discriminant score for a desired trait in each seed;
(4th step) The recovery rate and the conformity rate are defined by the following formula:
[Recovery rate (%)] = [Number of seeds selected based on the discriminant score among the seeds subjected to the quality evaluation] / [Number of seeds subjected to the quality evaluation] × 100
[Conformity rate (%)] = [Number of seeds having desired characteristics among the seeds subjected to the quality evaluation and selected based on the discriminant score] / [Number of seeds selected based on the discriminant score among the seeds subjected to the quality evaluation] × 100
In addition, when the selection conditions that make the preliminary non-defective rate and the recovery rate equal are defined as "standard selection conditions",
a step of selecting a discrimination model from the candidates of the quality discrimination model derived in the third step, the step including selecting, as a quality discrimination model, a candidate that satisfies a criterion that the matching rate exceeds a preliminary quality rate under the standard screening conditions;
(Step 5) A step of irradiating the remaining seeds not removed from the seed population in Step 1 and/or other seeds obtained under substantially the same conditions as those of the seeds with light to obtain spectral data, and applying the discrimination model selected in Step 4 to the spectral data to determine a discrimination score for each seed;
(Sixth step) determining a threshold value of the discriminant score based on the discriminant score determined in the fifth step;
(Step 7) A step of recovering seeds predicted to have desired traits by comparing the discrimination scores of each seed determined in step 5 with the discrimination score threshold determined in step 6. The selection method according to claim 1, which is a method for selecting plant seeds having two or more desired traits. The selection method according to claim 1 or 2, wherein the trait is at least one trait selected from the group consisting of germination ability, germination vigor, genotype, stress resistance, dormancy, disease resistance, insect resistance, QTL characteristics, eating quality, heading time, and morphological characteristics. The method according to claim 1 or 2, wherein the prediction of two or more desired traits is carried out by any one of the following methods (i) to (iii):
(i) A method of scoring seeds individually using a discrimination model for each trait, focusing on the quality of each trait, and predicting that seeds with all scores equal to or above a threshold are good seeds;
(ii) A method of selecting a single discrimination model by determining that a seed is good if all of the desired traits are good, and determining that the other traits are bad, regardless of whether each trait is good or bad, scoring the seeds using the single discrimination model, and predicting that seeds with a score equal to or higher than a threshold are good seeds;
(iii) A method in which a discrimination model for each trait is selected by focusing on the quality of each trait, the scores obtained using the discrimination model for each trait are integrated to obtain an integrated score, and the quality of each seed is predicted based on the ranking of the integrated score. The selection method according to claim 1 or 2, further comprising a step of evaluating the accuracy of the discrimination model. The selection method according to claim 1 or 2, wherein the plant is a horticultural crop.