Machine learning-based whole-genome selective breeding method for perfume coconut aroma substance 2AP

Abstract

The invention belongs to the field of coconut breeding, and particularly relates to a perfumed coconut aroma substance 2AP whole genome selective breeding method based on machine learning. The invention provides a perfume coconut 2AP content prediction method based on whole genome variation sites and a machine learning algorithm, and finally, 21 molecular marker sites are screened out as key features to construct a 2AP content prediction model for early germplasm screening. According to the method, efficient modeling of complex aroma traits is achieved, a reliable tool is provided for early screening of high-quality germplasm of Xianja, a key basis is provided for construction of a molecular prediction model, the screening efficiency is greatly improved, the cost is reduced, the aroma potential of filial generations can be predicted in the seedling stage, and the method is suitable for popularization and application. Therefore, the breeding period is effectively shortened, and cultivation and popularization of high-quality varieties of the perfume coconuts are accelerated.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of coconut breeding, and specifically relates to a method for breeding coconut varieties with 2AP based on machine learning. BACKGROUND

[0002] Coconut is an important economic crop in the tropics, and plays an important role in improving the nutritional supply of residents and promoting the income of farmers. Coconut can be divided into two main groups: high-type and dwarf-type coconut. Dwarf coconut is often used for fresh consumption due to its early flowering and high yield. Among them, the coconut variety with unique "mottled leaf fragrance" or "popcorn fragrance" in its coconut water and flesh is favored by consumers and its market price can be several times that of ordinary coconut. Studies have shown that the characteristic fragrance of coconut with unique fragrance mainly comes from 2-acetyl-1-pyrroline (2AP), which is a key metabolite that determines the fragrance of crops such as rice, soybeans, cucumbers, sorghum, and winter melons. In coconut, high levels of 2AP are mainly detected in the juice and endosperm of coconut with unique fragrance, while it is basically not detected in non-fragrant coconut. Further studies have found that the biosynthesis of 2AP is closely related to amino acid metabolism and polyamine metabolism, and its genetic regulation mechanism is mainly related to the loss or reduction of function of amino aldehyde dehydrogenase (AMADH) family genes. In coconut, studies have identified aromatic allelic variations of CnAMADH2, such as the single nucleotide mutation (G mutation to C) found in Thai green dwarf coconut, which causes the 422nd amino acid of the protein to change from alanine to proline, resulting in a reduction or loss of enzyme function, thereby promoting the accumulation of 2AP. Based on this, molecular markers have been developed to distinguish between fragrant and non-fragrant materials and applied to germplasm purification and molecular detection. However, studies have also found that some fragrant coconut varieties (such as yellow dwarf coconut) cannot be distinguished by a single marker, so there may be multiple allelic variations of coconut 2AP traits, and the genetic mechanism is complex.

[0003] Current research results show that 2AP content in coconut with unique fragrance is a typical quantitative trait, which is regulated by multiple genes and complex metabolic pathways. Traditional molecular detection methods relying on a single gene or a single marker have limited predictive effect in large-scale natural populations and diverse germplasm, making it difficult to support systematic molecular breeding needs. With the development of high-throughput sequencing and population genetics, genetic variations at the whole genome level have been widely applied to the study of crop flavor and quality traits. SUMMARY

[0004] The present application provides a method for breeding coconut varieties with 2AP based on machine learning.

[0005] The technical solution of the present application is implemented as follows:

[0006] The method for predicting the content of perfume coconut aroma substance 2AP based on machine learning comprises the following steps:

[0007] (1) Collect fresh coconut fruits to obtain 2AP content phenotype data; meanwhile, collect fresh coconut leaves to extract genomic DNA and perform whole-genome paired-end sequencing to obtain raw sequencing data;

[0008] (2) The raw sequencing data is subjected to quality control by Fastp (v0.23.2) software, low-quality sequencing reads, adapter contamination and errors in the sequence are removed, and then the cleaned data is aligned to the reference genome GCA_008124465.1_ASM812446v1 of Cocos nucifera by using BWA (Burrows-Wheeler Aligner) software; the generated SAM file is arranged by Samtools, converted into BAM format, and the repeated reads are removed by using Picard tool to ensure the accuracy of the data; next, the HaplotypeCaller tool in Genome Analysis Toolkit (GATK) is used to detect mutations in the BAM file, and a GVCF format file is generated; the GVCF files of each sample are combined by CombineGVCF software to generate a single combined GVCF file, and then the genotype data is extracted by GenotypeGVCF software; all variant data is subjected to further quality control and filtration by using PLINK software;

[0009] (3) Analyze the correlation between the variant data and the trait data to obtain a set of variant sites significantly related to 2AP content;

[0010] (4) The significantly related sites are sorted according to the -log10(P) value, and different numbers of sites are extracted as feature variables; different numbers of feature sites are input into a linear regression model for modeling;

[0011] (5) The best feature variable is finally selected by observing the influence of different numbers of sites on the model performance;

[0012] (6) After deleting the linked sites, the SNP sites used for model construction are finally selected, and a prediction model is constructed, as follows:

[0013] ;

[0014] (7) The constructed prediction model is used to predict the 2AP content of the samples to be tested.

[0015] Further, in the step (1), the determination of 2AP content is performed by using an Agilent 7890B gas chromatography-mass spectrometry (GC-MS) system.

[0016] Furthermore, in step (2), the method for controlling and filtering variant data is as follows: firstly, variant sites with a minimum allele frequency (MAF) of less than 0.005 are removed, and sites with a genotype deletion rate of greater than 0.5 are removed, while samples with a genotype quality (GQ) of greater than or equal to 20 are retained.

[0017] Furthermore, in step (3), the correlation between variation data and trait data is analyzed using a mixed linear model of GEMMA and the Wald p-value test, and GWAS analysis is performed.

[0018] Furthermore, in step (4), the number of feature sites extracted is the first 10, the first 100, the first 1000, the first 5000, and the first 10000.

[0019] Furthermore, in step (6), 21 sites were ultimately selected for model construction; the SNP sites are 1_23616816, 1_75608059, 1_95044121, 2_4481809, 3_31040303, 3_56688579, 5_1895872, 7_40953268, 7_62820209, 8_ β of 48206190, 8_48980279, 8_49173032, 8_49175687, 8_51337660, 8_51544209, 10_25132846, 10_33527500, 10_33527549, 11_10062240, 14_15152469, 14_42400666 j The values ​​are, in order: 0.3950234, 0.180863756, 0.282207482, -0.743907591, -0.634554563, -0.12887849, -0.092796702, 0.134043125, -0.342901713, -0.016155737, -0.016543865, and 0.003. 907421, 0.115235923, -0.069205112, 0.034534505, -0.176058524, -0.329789353, -0.385235499, -0.290224826, -0.102017657, 0.180096191, β0 value is -0.0865852515065637.

[0020] A set of molecular markers for predicting the 2AP content of perfume coconut includes the following SNP sites:

[0021]

[0022] In the above table, the "Sequence" is the nucleotide sequence of the SNP molecular marker; the position of the SNP molecular marker site is marked in "[]", and the base is the polymorphism of the site.

[0023] A screening method of perfume coconut germplasm, comprising the following steps:

[0024] (1) Determining the genotype of the SNP site in the above-mentioned molecular marker combination in the sample to be tested, and using the model constructed in the above-mentioned perfume coconut aroma substance 2AP content prediction method to predict the aroma substance 2AP content of the perfume coconut germplasm;

[0025] (2) Screening the perfume coconut germplasm according to the predicted aroma substance 2AP content of the perfume coconut.

[0026] A machine learning-based perfume coconut aroma substance 2AP whole genome selection breeding method, comprising the following steps:

[0027] (1) Determining the genotype of the SNP site in the above-mentioned molecular marker combination in the sample to be tested, and using the model constructed in the above-mentioned perfume coconut aroma substance 2AP content prediction method to predict the aroma substance 2AP content of the perfume coconut germplasm;

[0028] (2) Selecting suitable perfume coconut germplasm for variety breeding according to the predicted aroma substance 2AP content of the perfume coconut.

[0029] The above-mentioned prediction method, screening method and / or molecular marker combination are applied to perfume coconut breeding.

[0030] The beneficial effects of the present application are:

[0031] The present application proposes a perfume coconut 2AP content prediction method based on whole genome variation sites and machine learning algorithm, and finally selects 21 molecular marker sites as key features. By comparing in various regression models, the superiority of the linear regression model in the prediction of 2AP quantitative traits is verified. Under the condition of the first 70000 sites, the overall fitting effect of the model is best; when it is partially simplified to 10000 sites, highly accurate prediction results can still be obtained.

[0032] The method has good stability and accuracy in the verification of the training set and the test set. When 10,000 sites are used as characteristic variables, the Pearson correlation coefficient between the predicted value and the measured 2AP content is as high as 0.93, indicating that the model can accurately reflect the aroma level of the germplasm. In order to ensure the accuracy of the model, 21 sites are finally selected to construct the model, and the prediction accuracy of the model can reach 0.63, which can explain about 40% of the genetic variation, indicating that these markers play a key role in the prediction model and are the main effective sites for screening aroma variation, and can be used for early germplasm screening. In the breeding process of the perfume coconut, the aroma potential is predicted to help breeders select potential high-quality materials at the seedling stage, saving the cost of subsequent experimental detection and field evaluation. Compared with the traditional method relying on experimental detection, the present application realizes efficient modeling of complex aroma traits, provides a reliable tool for early screening of high-quality germplasm of perfume coconut, and provides a key foundation for the construction of a molecular prediction model, which not only greatly improves the screening efficiency and reduces the cost, but also can predict the aroma potential of hybrid offspring at the seedling stage, thereby effectively shortening the breeding cycle and accelerating the cultivation and popularization of high-quality varieties of perfume coconut. BRIEF DESCRIPTION OF DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0034] Figure 1 . 168 samples of perfume coconut 2AP content whole genome association analysis results, variation data set contains 29774228 single nucleotide polymorphism variation SNP sites.

[0035] Figure 2 . Whole genome selection flowchart.

[0036] Figure 3 . Comparison of prediction performance of different models on the training set.

[0037] Figure 4 . PLSRegression model whole genome selection verification set results. DETAILED DESCRIPTION

[0038] With reference to the accompanying drawings on which some embodiments of the application are shown, the principles of the application will be described. Obviously, the described embodiments are only some, but not all, of the embodiments of the application. Based on these embodiments of the application, all other embodiments obtained by those of ordinary skill in the art without creative efforts should fall into the scope of the protection of the application.

[0039] Example 1 Genome-wide association analysis and sample partitioning

[0040] Firstly, 168 germplasm resources of Cocos nucifera were collected from the Cocos germplasm garden of the Chinese Academy of Tropical Agricultural Sciences, and the 2AP content in the coconut meat was determined. The 2AP content was determined by Agilent 7890B gas chromatography-mass spectrometry (GC-MS) system to ensure accurate detection of aroma components. The specific operation steps are as follows: after collecting fresh fruits, the samples were first rapidly frozen in liquid nitrogen, and then ground into fine powder. 2 grams of coconut meat powder sample was added to 2 mL of anhydrous ethanol, shaken well and then heated in a 80°C water bath for 4 hours. After the sample cooled to room temperature, the supernatant was separated by centrifugation, and then transferred to a 1.5 mL sample bottle for direct GC-MS analysis. Mass spectrometry analysis was performed by comparing with the standard mass spectrometry database in NIST14 mass spectrometry database and the retention index (RI) information for qualitative and quantitative analysis, so as to obtain accurate 2AP content phenotype data. At the same time, the corresponding fresh coconut leaves were collected for extraction of genomic DNA, and the CTAB method was used for standardized DNA extraction to ensure that the extracted DNA quality was suitable for subsequent analysis. The extracted DNA was determined for concentration by Nanodrop spectrophotometer, and its quality was checked by 1% agarose gel electrophoresis. After the concentration of the extracted DNA was qualified, it was diluted to 100 ng / μL and stored at -20°C for subsequent library construction and sequencing.

[0041] The DNBSEQ platform was used for whole-genome paired-end sequencing, with a sequencing length of 150-bp, ensuring a depth of more than 15X for each sample to obtain sufficient genome coverage. The raw sequencing data was subjected to quality control using Fastp (v0.23.2) software to remove low-quality sequencing reads, adapter contamination, and errors in the sequence. Then, the cleaned data was aligned to the reference genome of Hyphaene thebaica (GCA_008124465.1_ASM812446v1) using BWA (Burrows-Wheeler Aligner) software. The SAM file generated after alignment was processed by Samtools to convert it to BAM format, and Picard tools were used to remove duplicate reads to ensure data accuracy. Next, the HaplotypeCaller tool in GATK (Genome Analysis Toolkit) was used to detect mutations in the BAM file, generating a GVCF format file. The GVCF files of each sample were combined to generate a single merged GVCF file using the CombineGVCF software, and then genotype data was extracted using the GenotypeGVCF software. All variant data was further quality controlled and filtered using PLINK software. First, variant sites with a minimum allele frequency (MAF) less than 0.005 were removed, while sites with a genotype missing rate greater than 0.5 were excluded, and samples with a genotype quality (GQ) greater than or equal to 20 were retained. After strict quality control and variant data processing, we finally obtained a high-quality whole-genome variant dataset, laying the foundation for subsequent genome-wide association analysis (GWAS).

[0042] On this basis, the mixed linear model of GEMMA and Wald P-value test were used to analyze the correlation between variant data and trait data, and GWAS analysis was performed to obtain a set of variant sites significantly associated with 2AP content. To further screen reliable markers associated with 2AP content, the GWAS results were sorted according to P-value, and the MAF of each site was calculated. Subsequently, the 168 samples were randomly divided into a training set (n=134) and a test set (n=34) in a 4:1 ratio for training and prediction evaluation of the machine learning model, respectively.

[0043] Example 2 Training and screening of a Hyphaene thebaica 2AP content prediction model

[0044] According to the results of GWAS analysis, we ranked the significant associated loci by -log10(P) value and extracted different number of loci (e.g. top 10, top 100, top 1000, top 5000, top 10000, etc.) as feature variables. We input different number of feature loci into 7 classical machine learning regression models for modeling and prediction, including Ridge, Lasso, Linear, KernelRidge, PLSRegression, ElasticNetCV and SVR_linear.

[0045] We compared the prediction performance of different models on the training set by cross-validation Figure 3 ), and the results showed that PLSRegression model performed better in stability and prediction correlation when the number of features was small, so Figure 4 We used PLSRegression model as an example to show the trend of prediction performance with the number of features, and by observing the influence of different number of loci on model performance, we finally selected 30 loci as the best feature variables. After deleting the linked loci, we finally selected 21 SNP markers for model construction (Table 1). The results of other models with different number of features fluctuated greatly, and the stability and prediction correlation were poor, and the prediction effect was significantly reduced.

[0046] Table 1: Information of 21 SNP markers

[0047]

[0048] In the above table, "Sequence" is the nucleotide sequence of the SNP marker; "[]" marks the position of the SNP marker site, and the base in it is the polymorphism of the site.

[0049] The constructed model is as follows:

[0050]

[0051] where, is the predicted 2AP content, SNP j is the genotype encoding of each SNP site, β j As shown in Table 2, the value of β0 is -0.0865852515065637.

[0052] Table 2: β coefficients of each SNP site

[0053]

[0054] Example 3 Phenotype prediction and evaluation of 2AP content in perfume coconut

[0055] Using the best model and 21 SNP loci screened, we predicted the 2AP content of the test set samples. The genotypes of the 21 SNP loci in the test set samples were detected, the genotype codes were obtained, and the 2AP content of each sample was predicted according to the constructed model.

[0056] The results show that when 10,000 loci are used, the Pearson correlation coefficient between the predicted value and the actual measured value can reach 0.93. This indicates that the established model can effectively replace part of the experimental detection and has the ability to predict the aroma potential of coconut hybrid offspring in the early seedling stage, thereby improving the efficiency of germplasm screening and accelerating the breeding process of high-quality varieties of perfume coconut.

[0057] The Pearson correlation coefficient in the verification using 21 molecular markers can reach 0.63, indicating that the model still has considerable prediction ability and can effectively reflect the genetic signal of 2AP content. Since the factors affecting 2AP content in perfume coconut are very complex, 21 SNP markers can reach 0.63, which means that they can explain about 40% of the genetic variation and can be used for early-stage germplasm screening. By predicting the aroma potential, breeders can select potential high-quality materials at the seedling stage, saving the cost of subsequent experimental detection and field evaluation.

[0058] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A machine learning-based method for predicting the content of 2AP, a fragrance compound in coconut perfumes, characterized in that... Includes the following steps: (1) Fresh coconut fruits were collected to obtain 2AP content phenotypic data; at the same time, fresh coconut leaves were collected for genomic DNA extraction, and whole genome paired-end sequencing was performed to obtain raw sequencing data; (2) The raw sequencing data were quality controlled using Fastp (v0.23.2) software to remove low-quality sequencing reads, adapter contamination and errors in the sequence. Then, the cleaned data was aligned to the reference genome of the fragrant coconut GCA_008124465.1_ASM812446v1 using BWA (Burrows-Wheeler Aligner) software. The SAM file generated after alignment was cleaned using Samtools and converted to BAM format. The Picard tool was used to remove duplicate reads to ensure the accuracy of the data. Next, the HaplotypeCaller tool in Genome Analysis Toolkit (GATK) was used to perform mutation detection on the BAM file and generate a GVCF format file. The GVCF files of each sample were merged using CombineGVCF software to generate a single merged GVCF file, and then genotype data were extracted using GenotypeGVCF software; all variation data were further quality controlled and filtered using PLINK software. (3) Analyze the correlation between variation data and trait data to obtain a dataset of variation sites that are significantly related to 2AP content; (4) The significantly correlated sites were sorted according to the -log10(P) value, and different numbers of sites were extracted as feature variables; the different numbers of feature sites were input into the linear regression model for modeling. (5) By observing the impact of different numbers of loci on the model performance, the best feature variables were finally selected; (6) After removing linkage sites, the final SNP sites selected for model construction are used to build the prediction model, as follows: ； (7) The constructed prediction model is used to predict the 2AP content of the sample to be tested.

2. The prediction method as described in claim 1, characterized in that, In step (1), the 2AP content was determined using an Agilent 7890B gas chromatography-mass spectrometry (GC-MS) system.

3. The prediction method as described in claim 1, characterized in that, In step (2), the method for controlling and filtering variant data is as follows: first, variant sites with a minimum allele frequency (MAF) of less than 0.005 are removed, and sites with a genotype deletion rate of greater than 0.5 are removed, while samples with a genotype quality (GQ) of greater than or equal to 20 are retained.

4. The prediction method as described in claim 1, characterized in that, In step (3), the correlation between variation data and trait data is analyzed using a mixed linear model of GEMMA and the Wald p-value test, and GWAS analysis is performed.

5. The prediction method as described in claim 1, characterized in that, In step (4), the number of feature sites extracted is the first 10, the first 100, the first 1000, the first 5000, and the first 10000.

6. The prediction method as described in claim 1, characterized in that, In step (6), 21 SNP sites were finally selected for model construction; the SNP sites are 1_23616816, 1_75608059, 1_95044121, 2_4481809, 3_31040303, 3_56688579, 5_1895872, 7_40953268, 7_62820209, 8_4 β of 8206190, 8_48980279, 8_49173032, 8_49175687, 8_51337660, 8_51544209, 10_25132846, 10_33527500, 10_33527549, 11_10062240, 14_15152469, 14_42400666 j The values ​​are, in order: 0.3950234, 0.180863756, 0.282207482, -0.743907591, -0.634554563, -0.12887849, -0.092796702, 0.134043125, -0.342901713, -0.016155737, -0.016543865, and 0.

003. 907421, 0.115235923, -0.069205112, 0.034534505, -0.176058524, -0.329789353, -0.385235499, -0.290224826, -0.102017657, 0.180096191, β0 value is -0.0865852515065637.

7. A set of molecular markers for predicting the 2AP content of perfume coconut, characterized in that, Including the following SNP sites: In the table above, "Sequence" represents the nucleotide sequence of the SNP molecular marker; "[]" indicates the location of the SNP molecular marker site, and the bases in the [] represent the polymorphism of that site.

8. A method for screening perfume coconut germplasm, characterized in that, Includes the following steps: (1) Determine the genotype of the SNP site in the molecular marker combination described in claim 7 in the sample to be tested, and use the model constructed in the prediction method described in any one of claims 1-6 to predict the content of aroma substance 2AP in the perfume coconut germplasm; (2) Based on the predicted content of 2AP, the germplasm of fragrant coconut was screened.

9. A machine learning-based genome-wide selection breeding method for 2AP, a fragrance compound in coconut perfumes, characterized in that... Includes the following steps: (1) Determine the genotype of the SNP site in the molecular marker combination described in claim 7 in the sample to be tested, and use the model constructed in the prediction method described in any one of claims 1-6 to predict the content of aroma substance 2AP in the perfume coconut germplasm; (2) Select suitable perfume coconut germplasm varieties for breeding based on the predicted content of 2AP aroma substances in perfume coconut.

10. The application of the prediction method as described in any one of claims 1-6, the screening method as described in claim 7, or the molecular marker combination as described in claim 9 in the breeding of fragrant coconut.

Images