A method for early prediction of core morphological characteristics of rhododendron leaf

Abstract

The present application relates to a kind of early prediction method of rhododendron leaf core morphological characteristics, belong to plant molecular breeding and forest genetic improvement technical field, the method includes the following steps: S1, collection and calculation leaf related phenotype data;S2, obtain the whole genome variation set containing population allele frequency information;S3, extract the core SNP site in whole genome range independent of each other, and convert genotype data into numerical type feature map;S4, combine the leaf related phenotype data of S1 with the genotype feature map of S3 to carry out whole genome prediction model;S5, calculate the genome estimated breeding value (GEBV) of the specific leaf feature of target seedling, and according to the breeding value, carry out assisted selection.Solve the long cycle of traditional phenotype selection and the multiple collinearity noise caused by excessive variation sites in non-model woody plant species, leading to model overfitting, difficult to take into account the efficient prediction of major and micro gene effects.

Description

Technical Field

[0001] This invention belongs to the field of plant molecular breeding and forest genetic improvement technology, and relates to an early prediction method for core morphological characteristics of rhododendron leaves. Background Technology

[0002] Rhododendrons, often hailed as the "King of Woody Flowers," rely heavily on leaf morphology, which is a key indicator of their overall ornamental value, plant structure, and environmental adaptability. Leaf width, petiole width, and their length-to-width ratio directly influence leaf spatial distribution and light-receiving area, and are also closely related to physiological functions such as drought resistance and photosynthetic efficiency. Therefore, targeted improvement of leaf morphology is a crucial goal in the breeding of new rhododendron varieties.

[0003] Traditional rhododendron breeding relies primarily on field phenotypic observations. However, rhododendrons have a long juvenile period, and quantitative traits such as leaf morphology are highly susceptible to interference from environmental factors (G×E interactions) such as light, temperature, and water during the seedling stage. This results in low accuracy of early screening based on the seedling stage, severely limiting breeding efficiency. In recent years, genome-wide association analysis (GWAS) has been used in plant molecular breeding to screen major variable sites by setting significance thresholds. However, rhododendron leaf morphology is a typical complex quantitative trait, controlled by a large number of "minor effects, multiple genes." Traditional GWAS misses a large number of minor genes due to its stringent significance thresholds, leading to "lack of heritability" and making it difficult to achieve high-precision phenotypic prediction in actual breeding. Genome-wide selection (GS) technology estimates the sum of the effects of all markers across the entire genome without the need for preset significance thresholds, theoretically capturing all genetic variation contributions. However, in non-model woody plants, the massive number of genome-wide variation sites introduces severe multicollinearity noise, leading to model overfitting. There is an urgent need in this field for a novel molecular breeding method that does not rely on GWAS significance, can effectively reduce dimensionality, and accurately capture subtle multi-gene effects, so as to achieve high-precision early prediction of rhododendron leaf morphology. Summary of the Invention

[0004] In view of this, the present invention provides an early prediction method for the core morphological characteristics of rhododendron leaves, which solves the problems of long phenotypic selection cycles and multicollinearity noise caused by too many variation sites in non-model woody plant species, which makes it difficult to achieve efficient prediction of both major and minor gene effects due to model overfitting.

[0005] The present invention solves the above-mentioned technical problems by means of the following solution:

[0006] A method for early prediction of core morphological features of rhododendron leaves, the method comprising the following steps:

[0007] S1. Obtain a natural population of rhododendrons as a training population sample, cultivate them in a consistent environment, collect and calculate leaf-related phenotypic data, including leaf width, leaf length-to-width ratio, petiole width, and petiole length-to-width ratio for each sample.

[0008] S2, high-throughput whole-genome sequencing of the sample is performed and aligned to the reference genome. After generating the original variants through single-sample variant detection, joint gene typing is performed to obtain a whole-genome variant set containing population allele frequency information.

[0009] S3, perform quality control on the whole genome variation set containing population allele frequency information, filter the minor allele frequency and genotype deletion rate, then filter out collinear variation sites based on the physical linkage disequilibrium within the sliding window, extract mutually independent core SNP sites across the whole genome, and convert the genotype data into numerical feature maps;

[0010] S4, combining the leaf-related phenotypic data of S1 with the genotypic feature map of S3, using the whole genome regression algorithm, whole genome prediction models for leaf width, leaf length-to-width ratio, petiole width and petiole length-to-width ratio are respectively.

[0011] S5. Extract the genotype data of the rhododendron seedlings to be screened at the core SNP loci, input them into the prediction model constructed in S4, calculate the genomic estimated breeding value (GEBV) of specific leaf characteristics of the target seedlings, and perform auxiliary selection based on the breeding value.

[0012] Furthermore, in S4, the whole genome regression algorithm is selected from one of the following: Reproducible Kernel Hilbert Space (RKHS) model, Ridge Regression Best Linear Unbiased Prediction (RR-BLUP) model, BayesC model, or BayesL model.

[0013] Furthermore, when predicting the leaf length-to-width ratio of phenotypic data, the RR-BLUP or RKHS model is preferred; when predicting the leaf width of phenotypic data, the RKHS model is preferred; when predicting the petiole width of phenotypic data, the BayesC model or RR-BLUP model is preferred; and when predicting the petiole length-to-width ratio of phenotypic data, the BayesL model is preferred.

[0014] Furthermore, when the whole-genome regression algorithm is the RKHS model, the squared Euclidean distance between samples is calculated based on the numerical feature matrix, and a nonlinear Gaussian kernel matrix is ​​constructed as the covariance matrix to capture the additive and epistatic effects of the marker set on leaf morphological features.

[0015] Furthermore, the reference genome in S2 is a Rhododendron genome sequence.

[0016] Furthermore, the genome sequence is preferably the genome sequence of *Rhododendron platypodum*. Specifically, the genome sequence of *Rhododendron platypodum* is publicly available at the National Genome Science Data Center, with BioProject number PRJCA038553 and GSA original sequence number CRR1762019.

[0017] Furthermore, the preferred operation for filtering minor allele frequency and genotype deletion rate in S3 is to remove sites with a deletion rate > 10% and a minor allele frequency < 0.05.

[0018] Furthermore, the preferred operation in S3 for filtering out collinear variant sites by setting physical linkage disequilibrium within a sliding window is as follows: within a 50 kb physical range, remove r 2 Strong collinearity markers exceeding 0.2.

[0019] After performing the above filtering on allele frequencies and genotype deletion rates, and then filtering out collinear variant sites based on the physical linkage disequilibrium within the sliding window, 15,160 independent core SNP sites were extracted across the entire genome.

[0020] Beneficial effects:

[0021] 1. This invention is the first to apply whole-genome selection technology to the early prediction of leaf morphology characteristics in rhododendrons. Only low-depth sequencing of seedling rhododendrons and extraction of core SNP markers are required. The model can accurately predict the leaf type of mature plants and eliminate inferior individual plants. This completely changes the passive situation in traditional breeding where the plant must be grown to maturity before leaf morphology identification can be carried out. It greatly shortens the breeding cycle and reduces the cost of land and artificial cultivation.

[0022] 2. Leaf morphology in non-model woody plants like rhododendrons is a complex quantitative trait controlled by a large number of small-effect multi-genes. This invention employs a labeling and dimensionality reduction strategy that does not rely on any phenotypic association statistical thresholds, effectively eliminating genetic multicollinearity noise and achieving differentiated and accurate modeling of leaf traits with different genetic structures. The predicted correlation coefficient for leaf length-to-width ratio reached 0.85, the correlation coefficient for petiole width reached 0.820, the coefficient for leaf width reached 0.749, and the predicted coefficient for petiole length-to-width ratio reached 0.677, significantly higher than the 0.251 predicted by the full variation map. Furthermore, the predicted scatter points converged tightly to the diagonal ideal prediction line, indicating that the model is stable and reliable and can be directly used for industrial breeding decision support. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the technical process;

[0024] Figure 2 This is a scatter plot showing the linear fit between the actual observed values ​​of core leaf features in the independent validation population and the genome-estimated breeding value (GEBV) in Example 1. Detailed Implementation

[0025] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.

[0026] Example 1:

[0027] A method for early prediction of core morphological features of rhododendron leaves, flowchart as follows: Figure 1 As shown, the specific steps are as follows:

[0028] S1, Experimental population and phenotypic determination.

[0029] Eighty-four representative evergreen rhododendron germplasm resources were selected to form an experimental population. These were strictly divided into two parts: 68 accessions served as a model training population, and 16 accessions served as a completely independent external validation population. Under consistent greenhouse cultivation conditions, the leaf width and petiole width of each individual plant were precisely measured using calipers. Combined with leaf length and petiole length, four core leaf morphological characteristics—leaf length-to-width ratio, petiole length-to-width ratio, leaf width, and petiole width—were calculated to obtain phenotypic data.

[0030] S2, sequencing and high-density variant detection.

[0031] High-quality genomic DNA was extracted from 84 samples and high-throughput whole-genome resequencing was performed using the Illumina platform. The raw data underwent rigorous quality control using FASTP software. The quality control filtering rules included: removing reads containing sequencing adapters, removing reads with an N base ratio greater than 10%, removing reads with a low-quality base ratio (Q ≤ 20) exceeding 50%, and discarding reads with a trimmed length < 50 bp. Clean reads that passed quality control were aligned to the Rhododendron reference genome using the BWA MEM algorithm.

[0032] To improve the efficiency and accuracy of variant detection in large populations, a parallel "scatter-gather" strategy was adopted. The HaplotypeCaller module of GATK software was used to extract variants from each sample concurrently on a chromosome basis, generating independent GVCF files. Subsequently, CombineGVCFs and GenotypeGVCFs were used to perform joint generative morphology detection, accurately capturing the confidence levels of allele frequencies and non-mutant background sites in the population, outputting a high-quality whole-genome SNP dataset containing 84 samples.

[0033] The reference gene for rhododendron is the Rhododendron platypodum gene, which is publicly available in the National Genome Science Data Center. Its BioProject number is PRJCA038553, and its original GSA sequence is CRR1762019.

[0034] S3, dimensionality reduction of the core feature matrix.

[0035] To eliminate high-dimensional genetic noise and avoid the significance threshold dependence of the GWAS method, this embodiment directly performs quality control and dimensionality reduction on the whole genome SNP set.

[0036] Low-quality loci (MAF) with a minor allele frequency (MAF) < 0.05 and a deletion rate greater than 10% were filtered out using PLINK software. The linkage disequilibrium coefficient r between markers was calculated within a 50 kb sliding window. 2 Remove r 2 Strong collinearity markers with a value > 0.2 were identified. Ultimately, 15,160 non-linked "golden" core SNPs distributed throughout the genome were retained and transformed into a numerical feature matrix (where 0, 1, and 2 represent homozygous reference, heterozygous, and homozygous mutant genotypes, respectively). For the very few missing values, mean imputation was performed using allele frequencies. The average genotype value of the effective genotypic samples for that specific marker in the entire population was then used to fill in the missing values, thus obtaining a complete and missing numerical genotype feature map.

[0037] S4-S5, Training and Independent Blind Testing of Bayesian and Nonlinear Kernel Models.

[0038] In the R language environment, the population leaf phenotypic data in S1 were combined with the numerical genotypic feature map in S3. To address the genetic complexity of different traits, seven whole genome regression models, namely GBLUP, Reproducible Kernel Hilbert Space (RKHS), RR-BLUP, BayesA, BayesB, BayesC, and BayesL, were constructed and trained in parallel. More than 3,000 MCMC iterations were performed to ensure model convergence, and genomic selection prediction models for rhododendron leaf phenotypic data were constructed respectively.

[0039] The RKHS model uses the Gaussian distance formula to map 15,160 SNP features into a nonlinear kernel matrix. After the model is trained, the leaf phenotypes of 16 external validation populations are "masked" (forced to be NA) and input into the model for blind guessing prediction to calculate their genome-estimated breeding value (GEBV).

[0040] The results are shown in Table 1 and Figure 2 As shown.

[0041] Table 1. Independent blind test accuracy of rhododendron leaf traits under the prediction model (Pearson's r)

[0042]

[0043] Results analysis:

[0044] Combined with Table 1 and Figure 2 The verification results show that the method provided by this invention exhibits extremely high prediction accuracy in independent verification populations, successfully overcoming the technical bottleneck of the inability to accurately predict complex quantitative traits at the seedling stage. Specifically:

[0045] Leaf length-to-width ratio: Under the RR-BLUP and RKHS models, the Pearson correlation coefficient (r) between GEBV and the actual phenotype is as high as 0.857 and 0.856, respectively;

[0046] Petiole width: Under the BayesC and RR-BLUP models, the r values ​​reached 0.820 and 0.810, respectively;

[0047] Ye Kuan: The prediction accuracy under the RKHS model reached 0.749;

[0048] Petiole length-to-width ratio: The BayesL model achieved a prediction accuracy of 0.677.

[0049] Depend on Figure 2As can be seen from the fitted scatter plot, the predicted scatter points converge tightly to the diagonal ideal prediction line, proving that the extracted core SNP matrix has fully captured the micro-effect multi-gene regulatory network that determines the leaf morphology of rhododendrons. In practical breeding applications, breeders only need to perform low-depth sequencing on seedling rhododendrons and extract the above-mentioned core SNPs to accurately predict their adult leaf type using this model and eliminate inferior individual plants.

[0050] Comparative experiment:

[0051] Current conventional techniques generally suggest that including as many variation sites in the genome as possible can maximize the capture of genetic variance. However, this study used a conventional unpruned full variance map to predict the core nutritional trait "leaf length-to-width ratio" and compared it with Example 1 of this invention in parallel cross-validation.

[0052] Similarly, 84 adult evergreen rhododendrons with different genetic backgrounds selected in Example 1 were used as research materials, and were strictly divided into 68 model training groups and 16 independent external validation groups (juvenile seedlings simulating unknown phenotypes).

[0053] Experimental Procedure: This study designed a rigorous parallel control experimental system for genome-wide selection prediction of leaf traits. The specific execution steps are as follows:

[0054] Step 1: Construction of Full and Dimensionally Reduced Datasets. First, collect leaf length-to-width ratio phenotypic data from natural populations. Extract the full set of 2.74 million unfiltered and unpruned SNPs to construct a high-dimensional genotype feature matrix, serving as a comparative dataset reflecting conventional techniques.

[0055] Using the specific dimensionality reduction method in step S3 of Example 1, 15160 core SNPs were extracted to construct a low-dimensional feature matrix without redundancy, which served as the dataset for Example 1.

[0056] Step 2: Cross-validation strategy setting. For the seven mainstream genome selection models (GBLUP, RKHS, RR-BLUP, BayesA, BayesB, BayesC, and BayesL), a uniform 30-fold independent cross-validation was set. The sample groups of the comparative dataset and the Example 1 dataset were strictly divided into a model training set (64 rhododendron leaf samples) and a blind test validation set (16 rhododendron samples) in the same proportion.

[0057] Step 3: Parallel Model Training and Computation. The comparative dataset matrix of 2.74 million and the Example 1 dataset matrix of 15,000 were independently input into the seven models mentioned above for fitting and computation, performing 3000 high-intensity MCMC iterations. During the computation, the computational resources and total time required for both sets of data to complete iterative verification of the same scale were systematically recorded.

[0058] Step 4: Accuracy Assessment and Results Summary. After the calculation, calculate the Pearson correlation coefficient between the predicted values ​​output by the two models and the actual leaf length-to-width ratio observations in the validation set. Summarize the time difference and accuracy difference to objectively assess the degree of interference of redundant noise on the advanced statistical model and the feasibility of industrial application.

[0059] The obtained data is shown in Table 2:

[0060] Table 2. Prediction accuracy of the "Comparative Example" model on the aspect ratio of rhododendron leaves.

[0061]

[0062] Results analysis:

[0063] Directly applying the full range of 2.74 million variants not only leads to a significant waste of computing resources, increasing model analysis time from one day to six days, but also causes a precipitous drop in the generalization accuracy of all models due to severe multicollinearity and the curse of dimensionality. Under the interference of massive amounts of invalid noise, the prediction accuracy of conventional statistical models such as GBLUP and RR-BLUP plummeted to an extremely low 0.233, rendering the models almost ineffective; even Bayesian models with variable selection capabilities hovered only around 0.24-0.25, completely losing their practical application value for early-stage auxiliary selection in breeding. This further demonstrates the necessity and significant progress of the core marker set extraction strategy of this invention in purifying genetic information and overcoming high-dimensional noise. Furthermore, it shows that the method provided by this invention exhibits extremely high prediction accuracy in independent validation populations, successfully overcoming the technical bottleneck of the inability to accurately predict complex quantitative traits at the seedling stage.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for early prediction of core morphological features of rhododendron leaves, characterized in that, The method includes the following steps: S1. Obtain a natural population of rhododendrons as a training population sample, cultivate them in a consistent environment, collect and calculate leaf-related phenotypic data, including leaf width, leaf length-to-width ratio, petiole width, and petiole length-to-width ratio for each sample. S2, high-throughput whole-genome sequencing of the sample is performed and aligned to the reference genome. After generating the original variants through single-sample variant detection, joint gene typing is performed to obtain a whole-genome variant set containing population allele frequency information. S3, perform quality control on the whole genome variation set containing population allele frequency information, filter the minor allele frequency and genotype deletion rate, then filter out collinear variation sites based on the physical linkage disequilibrium within the sliding window, extract mutually independent core SNP sites across the whole genome, and convert the genotype data into numerical feature maps; S4, combining the leaf-related phenotypic data of S1 with the genotypic feature map of S3, using the whole genome regression algorithm, whole genome prediction models for leaf width, leaf length-to-width ratio, petiole width and petiole length-to-width ratio are respectively. S5. Extract the genotype data of the rhododendron seedlings to be screened at the core SNP loci, input them into the prediction model constructed in S4, calculate the genomic estimated breeding value (GEBV) of specific leaf characteristics of the target seedlings, and perform auxiliary selection based on the breeding value.

2. The method according to claim 1, characterized in that, In S4, the whole genome regression algorithm is selected from one of the following: Reproducible Kernel Hilbert Space (RKHS) model, Ridge Regression Best Linear Unbiased Prediction (RR-BLUP) model, BayesC model, or BayesL model.

3. The method according to claim 2, characterized in that, When predicting the leaf length-to-width ratio of phenotypic data, the RR-BLUP or RKHS model is preferred; when predicting the leaf width of phenotypic data, the RKHS model is preferred; when predicting the petiole width of phenotypic data, the BayesC model or RR-BLUP model is preferred; when predicting the petiole length-to-width ratio of phenotypic data, the BayesL model is preferred.

4. The method according to claim 1, characterized in that, The reference genome in S2 is the genome sequence of the genus Rhododendron.

5. The method according to claim 4, characterized in that, The preferred genome sequence is that of Rhododendron platypodum.

6. The method according to claim 1, characterized in that, The preferred operation for filtering minor allele frequency and genotype deletion rate in S3 is to remove sites with a deletion rate > 10% and a minor allele frequency < 0.

05.

7. The method according to claim 6, characterized in that, The preferred operation in S3, which filters out collinear variant sites by setting physical linkage disequilibrium within a sliding window, is as follows: within a 50 kb physical range, remove r... 2 Strong collinearity markers exceeding 0.2.

Images