A machine learning-based tea tree fuzz phenotype prediction method and system

Abstract

The application belongs to the technical field of bioinformatics, single-cell multi-omics and artificial intelligence, and discloses a tea tree hair phenotype prediction method and system based on machine learning. The method comprises the following steps: firstly, the development trajectory of tea tree epidermis and hair cells is finely analyzed by using single-cell transcriptome sequencing (scRNA-seq), and hair-specific expression gene clusters are accurately positioned. Combined with a high-density genetic map, the hair-specific expression genes screened at the single-cell level are subjected to expression quantitative trait locus (eQTL) mapping, 111 core single nucleotide polymorphism (SNP) sites most significantly related to hair development are strictly screened from the whole genome, and a Core-111 feature set with extremely reduced dimension is constructed. Subsequently, a variety of regression algorithms are evaluated in parallel, and an arithmetic mean ensemble learning model with ridge regression and gradient boosting regression as base learners is constructed.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of bioinformatics, single-cell multi-omics and artificial intelligence, and specifically relates to a method and system for predicting the pubescent phenotype of tea trees based on machine learning. Background Technology

[0002] In tea trees ( Camellia sinensis In the breeding of perennial crops such as tea leaves, the density of leaf hairs directly determines key qualities such as the "downy aroma" of tea. Traditional breeding processes are lengthy and susceptible to interference from the GxE effect; therefore, predicting phenotypes based on genotypes is crucial for achieving precision breeding.

[0003] However, traditional quantitative trait locus (QTL) mapping and genome-wide selection (GS) methods are severely ineffective when dealing with small-scale populations in actual breeding (such as F1 generations with fewer than 100 individuals). Existing studies largely rely on bulk RNA-seq, a method highly susceptible to interference from complex cellular heterogeneity within tissues—specific regulatory signals from epidermal trichomes are often masked by background noise from a vast number of mesophyll cells. Directly inputting these redundant SNPs, containing numerous non-coding regions and lacking biological mechanism validation, into machine learning models inevitably leads to the "curse of dimensionality," resulting in high computational costs and severe model overfitting. Therefore, there is an urgent need to develop a prediction scheme based on single-cell precision screening of core regulatory sites combined with the advantages of ensemble learning. Summary of the Invention

[0004] To address the issues of feature redundancy and noise interference caused by cellular heterogeneity in predicting complex quantitative traits using traditional whole-tissue sequencing, this invention provides a machine learning-based method and system for predicting the pubescent phenotype of tea trees. It mainly achieves accurate phenotype prediction through four collaborative modules: data preprocessing, single-cell precision dimensionality reduction, integrated model construction, and engineered prediction.

[0005] To achieve the above objectives, the present invention provides the following solution: A machine learning-based method for predicting the pubescent phenotype of tea trees, the method comprising: Single-cell transcriptome sequencing (scRNA-seq) was performed on the test material to obtain cell type-specific genes; transcriptome sequencing was performed on the test population to construct a high-density genetic map and extract a set of single nucleotide polymorphism (SNP) markers. For candidate genes specifically expressed by trichome cells identified through single-cell sequencing and trichome phenotype screening, eQTL mapping of expression quantitative trait sites was performed to identify and extract associated SNP sites. Based on associated SNP sites, an ensemble learning model based on the fusion of multiple algorithms is constructed. By using the best ensemble learning model trained on the full dataset, the genotypes of the test samples are aligned in feature space to obtain the predicted values ​​of the tea tree hair phenotype.

[0006] Preferably, the method for performing eQTL mapping on candidate genes specifically expressed by trichome cells identified through single-cell sequencing and trichome phenotype screening, and for identifying and extracting associated SNP sites, includes: scRNA-seq was used to analyze the cellular heterogeneity of tea tree shoots, identify trichome cell subpopulations, and extract several core candidate genes specifically expressed in relation to the target phenotype. By combining high-density genetic maps, eQTL mapping of expression quantitative trait loci was performed on trichome-specific expression genes screened at the single-cell level. 111 core single nucleotide polymorphism (SNP) loci that are most significantly associated with trichome development were screened from the whole genome, and a dimensionality-reduced Core-111 feature set was constructed.

[0007] Preferably, the method for constructing a dimensionality-reduced Core-111 feature set by combining high-density genetic maps, mapping expression quantitative trait loci (eQTLs) of trichome-specific expression genes screened at the single-cell level, and screening 111 core single nucleotide polymorphism (SNP) sites most significantly associated with trichome development from the whole genome includes: During the data quality control phase, non-signal variant sites with Mendelian errors or a deletion rate exceeding 1 / 3 are removed. Within the trichome cell subset, the effect of genotypic variation on the single-cell expression level of target genes was assessed using one-way ANOVA. For a significant regulatory region corresponding to the same target gene, the significance levels of all SNPs within that region are compared. Only the single site with the smallest P-value is retained as the representative SNP. Finally, 111 single-cell core regulatory SNP sites with the most significant correlation are selected, and a Core-111 feature set with mechanistic explanation and extreme dimensionality reduction is constructed.

[0008] Preferably, the ensemble learning model achieves quantitative prediction of the tea plant hair phenotype by weighted ensemble of base learners including ridge regression and gradient boosting regression. Specifically: Quantitative prediction value for any sample feature X The calculation formula is: ; in, This represents the predicted output of the ridge regression model. This represents the predicted output of the gradient boosting regression model.

[0009] Preferably, the method for obtaining tea tree pubescence phenotypic prediction values ​​after aligning the genotypes of the test samples with the feature space using the best ensemble learning model trained on the full dataset includes: The original dataset is input into the selected best ensemble learning model for final training. The original dataset contains the genotype matrix X and phenotype vector Y for all known phenotypes. Specifically: when the best model name contains "+", it is first split into the corresponding set of base model names, and then each base model is trained using fit(X, Y) on the complete original dataset. For the best combined model Ridge+GB, the Ridge model and the GB model are trained on the complete original dataset respectively to obtain the final model for predicting new materials, represented as: ; in, and These represent the final Ridge and GB models trained on the complete original dataset, respectively. The genotype data of the new samples to be tested are obtained and transposed. A forced rearrangement algorithm is used to strictly align the feature columns of the new samples to the column space dimension of the training set. Specifically: after obtaining the genotype data of the new samples, the data is read in using the group as the index column; it is first transposed to ensure that the new sample matrix and the training matrix maintain the same orientation, i.e., samples as rows and SNP sites as columns; subsequently, the training data matrix... Using the column names as criteria, the transposed new sample matrix is ​​rearranged; any redundant sites in the new sample that do not belong to the training column space are discarded during the rearrangement process; this process is represented as: ; in, This represents the genotype matrix of the new sample after transposition. This represents the new sample input matrix after rearranging according to the column order of the training matrix. Indicates a column; The aligned matrix is ​​input into the ensemble learning model after iterative fitting of parameters, and the output prediction results are generated to produce a two-dimensional correlation report containing sample IDs and predicted phenotypic values. Specifically: the aligned new sample matrix is ​​input into the optimal combination model Ridge+GB, and the two sub-models Ridge and GB output prediction values ​​respectively. The two sets of prediction results are then averaged to obtain the final predicted phenotypic value of the new material. ; Subsequently, the new sample numbers and corresponding predicted values ​​were compiled into a two-dimensional results table for phenotypic prediction and ranking analysis of new materials, which can be directly used to assist molecular breeding decisions.

[0010] The present invention also provides a tea tree hair phenotype prediction system based on machine learning. The system is used to implement the aforementioned method and includes: a data acquisition module, a site screening module, a model building module, and a phenotype prediction module. The data acquisition module is used to perform single-cell transcriptome sequencing (scRNA-seq) on the test material to obtain cell type-specific genes; and to perform transcriptome sequencing on the test population to construct a high-density genetic map and extract a set of single nucleotide polymorphism (SNP) markers. The site screening module is used to perform eQTL mapping on candidate genes specifically expressed by trichome cells selected by single-cell sequencing and trichome phenotype, and to identify and extract associated SNP sites. The model building module is used to build an ensemble learning model based on the fusion of multiple algorithms based on associated SNP sites. The phenotypic prediction module is used to obtain tea tree hair phenotypic prediction values ​​by aligning the genotypes of the test samples in the feature space using the best ensemble learning model trained on the full dataset.

[0011] Preferably, the site screening module includes: a single-cell clustering unit and a feature dimensionality reduction unit; The single-cell clustering unit is used to analyze the heterogeneity of tea tree shoot cells using scRNA-seq, identify trichome cell subpopulations, and extract several specific expression core candidate genes related to the target phenotype. The feature reduction unit is used to combine a high-density genetic map, perform eQTL mapping on trichome-specific expression genes screened at the single-cell level, screen out 111 core single nucleotide polymorphism (SNP) sites most significantly associated with trichome development from the whole genome, and construct a dimensionality-reduced Core-111 feature set.

[0012] Preferably, combining high-density genetic maps, the process of mapping expression quantitative trait loci (eQTLs) of trichome-specific expression genes screened at the single-cell level, and screening 111 core single nucleotide polymorphism (SNP) sites most significantly associated with trichome development from the whole genome to construct a dimensionality-reduced Core-111 feature set includes: During the data quality control phase, non-signal variant sites with Mendelian errors or a deletion rate exceeding 1 / 3 are removed. Within the trichome cell subset, the effect of genotypic variation on the single-cell expression level of target genes was assessed using one-way ANOVA. For a significant regulatory region corresponding to the same target gene, the significance levels of all SNPs within that region are compared. Only the single site with the smallest P-value is retained as the representative SNP. Finally, 111 single-cell core regulatory SNP sites with the most significant correlation are selected, and a Core-111 feature set with mechanistic explanation and extreme dimensionality reduction is constructed.

[0013] Preferably, the ensemble learning model achieves quantitative prediction of the tea plant hair phenotype by weighted ensemble of base learners including ridge regression and gradient boosting regression. Specifically: Quantitative prediction value for any sample feature X The calculation formula is: ; in, This represents the predicted output of the ridge regression model. This represents the predicted output of the gradient boosting regression model.

[0014] Preferably, the phenotypic prediction module includes: a global parameter fitting unit, a feature rearrangement and alignment unit, and a result output unit; The global parameter fitting unit is used to input the original dataset into the selected best ensemble learning model for final training. The original dataset contains all genotype matrices X and phenotype vectors Y for known phenotypes. Specifically: when the best model name contains "+", it is first split into the corresponding set of base model names, and then each base model is called to perform fit(X, Y) training on the complete original dataset. For the best combined model Ridge+GB, that is, training the Ridge model and the GB model on the complete original dataset respectively, the final model for new material prediction is obtained, represented as: ; in, and These represent the final Ridge and GB models trained on the complete original dataset, respectively. The feature rearrangement and alignment unit is used to acquire and transpose the genotype data of the new sample to be tested. Through a forced rearrangement algorithm, the feature columns of the new sample are strictly aligned to the column space dimension of the training set. Specifically: after acquiring the genotype data of the new sample to be tested, the data is read in using the group as the index column; it is first transposed to ensure that the new sample matrix and the training matrix maintain the same orientation, i.e., samples as rows and SNP sites as columns; subsequently, the training data matrix... Using the column names as criteria, the transposed new sample matrix is ​​rearranged; any redundant sites in the new sample that do not belong to the training column space are discarded during the rearrangement process; this process is represented as: ; in, This represents the genotype matrix of the new sample after transposition. This represents the new sample input matrix after rearranging according to the column order of the training matrix. Indicates a column; The result output unit is used to input the aligned matrix into the ensemble learning model after iterative fitting of the parameters, output the prediction results, and generate a two-dimensional correlation report containing sample IDs and predicted phenotypic values. Specifically: the aligned new sample matrix is ​​input into the optimal combination model Ridge+GB, and the two sub-models Ridge and GB output the predicted values ​​respectively. The two sets of prediction results are then averaged to obtain the final predicted phenotypic value of the new material. ; Subsequently, the new sample numbers and corresponding predicted values ​​were compiled into a two-dimensional results table for phenotypic prediction and ranking analysis of new materials, which can be directly used to assist molecular breeding decisions.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: The purpose of this invention is to provide a machine learning-based method for predicting the pubescent phenotype of tea trees. (1) Achieving ultimate dimensionality reduction with single-cell precision: Single-cell transcriptome sequencing (scRNA-seq) is introduced to finely analyze the developmental trajectory of trichome cells, screen trichome-specific expression genes at the single-cell level, and then through expression quantitative trait loci (eQTL) mapping, the input features are precisely condensed from massive whole genomes to core loci (111 core SNPs) that can only cause changes in gene expression in target cells. This fundamentally eliminates the computational time consumption and noise interference caused by heterogeneous cell population variations.

[0016] (2) Integrating the advantages of heterogeneous ensemble learning: abandoning the single heavy network that is prone to overfitting, constructing an arithmetic mean ensemble model with ridge regression (to handle collinearity and additive effects) and gradient boosting regression (to handle nonlinear interactions) as base learners, breaking through the bottleneck of prediction accuracy and stability under extremely small sample conditions. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of a genetic map constructed based on SNP markers according to an embodiment of the present invention; Figure 2This is a schematic diagram of the expression profile of trichome-specific genes and some SNPs in an embodiment of the present invention. In this diagram, A is a bubble diagram of trichome-specific cell expression based on single-cell transcriptome analysis; B is a box plot of expression of different genotypes of a specific core SNP. Figure 3 This is a scatter plot showing the regression correlation between predicted and actual phenotypic values ​​of the optimal ensemble model (Ridge+GB) in this embodiment of the invention on the test set. Figure 4 This is a graph showing the contribution of core SNPs to phenotypes based on SHAP values ​​in an embodiment of the present invention. Figure 5 This is a schematic diagram of a machine learning-based tea tree pubescence phenotypic prediction method according to an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Example 1 like Figure 5 As shown, this invention provides a machine learning-based method for predicting the pubescent phenotype of tea trees. Specifically, it involves a quantitative prediction method for the pubescent phenotype of tea trees based on single-cell precision biological mechanism dimensionality reduction and machine learning model integration. This method addresses the feature redundancy problem caused by mixed tissue cellular heterogeneity interference and the inclusion of a large number of non-coding regions or unrelated variations in existing genome-wide predictions when directly predicting phenotypes in small sample populations (such as F1 generation populations with fewer than 100 cases). It also overcomes the technical deficiency of existing feature screening methods that lack single-cell precision explanations of biological regulatory mechanisms. Furthermore, it addresses the technical challenges of insufficient fitting ability, large prediction bias, and weak generalization ability of single machine learning algorithms when facing quantitative traits of crops controlled by complex, subtle multi-gene interactions. The method includes: S1: Multi-omics data acquisition and feature initialization. Single-cell transcriptome sequencing (scRNA-seq) was performed on the test materials 'Bai Mao 2' (hairy material) and 'Ku Wei Bai Mu Dan' (sparsely hairy material) to obtain cell type-specific genes. Transcriptome sequencing was performed on the test population to construct a high-density genetic map containing 15 linkage groups, and a set of non-redundant single nucleotide polymorphism (SNP) markers was extracted. S2: Single-cell precision core feature screening (eQTL mapping). For candidate genes specifically expressed by trichome cells identified through single-cell sequencing and trichome phenotype screening, expression quantitative trait loci (eQTL) mapping is performed to identify and extract significantly associated core SNP sites as the input feature set for machine learning; S3: Construction and Training of Ensemble Learning Predictive Model. An ensemble learning model based on multi-algorithm fusion is constructed. This model achieves quantitative prediction of trichome phenotypes by weighted ensemble of base learners, including Ridge Regression and Gradient Boosting Regressor. S4: Sample Phenotypic Prediction and Result Output. Using the best ensemble learning model trained on the full dataset, the genotypes of the test samples are aligned in feature space, and the predicted trichome phenotype values ​​are output.

[0022] In this embodiment, multi-omics data acquisition and genotypic feature initialization are performed: single-cell transcriptome (scRNA-seq) and population sequencing data of the test materials are acquired. In this embodiment, tea plant (…) is selected. Camelliasinensis The cultivars 'Gui Re 2' (GR2) and 'Jin Mudan' (JMD) and their 77 F1 generation populations were used as research subjects. To ensure the integrity and reliability of the input model data, the original VCF genotype file containing 79 samples was first filled with missing values ​​and corrected for heterozygous sites using Beagle software based on a hidden Markov model. The effective threads nthreads=8 were specified in the running command parameters to accelerate the computation.

[0023] Based on high-quality sequencing data, a high-density genetic map was constructed using LEPMAP3 software. The map constructed in this embodiment contains 15 linkage groups (corresponding to Chr1-Chr15 in tea plants), with a total linkage group length of 1258.5 cM. After integrating 92,497 original SNP markers, the map underwent rigorous filtering to remove redundancy, ultimately retaining 1,785 non-redundant SNP markers as the initial feature set. The average map distance was 0.71 cM / marker, providing a high-quality genetic foundation for subsequent fine mapping and machine learning.

[0024] In this embodiment, the method for performing eQTL mapping on candidate genes specifically expressed by trichome cells, selected through single-cell sequencing and trichome phenotype screening, and identifying and extracting associated SNP sites includes: scRNA-seq was used to analyze the cellular heterogeneity of tea tree shoots, identify trichome cell subpopulations, and extract several core candidate genes specifically expressed in relation to the target phenotype. By combining high-density genetic maps, eQTL mapping of expression quantitative trait loci was performed on trichome-specific expression genes screened at the single-cell level. 111 core single nucleotide polymorphism (SNP) loci that are most significantly associated with trichome development were screened from the whole genome, and a dimensionality-reduced Core-111 feature set was constructed.

[0025] Specifically, traditional whole-genome prediction directly uses massive amounts of SNP or bulk RNA-seq data, which easily introduces cellular heterogeneity noise, leading to feature redundancy and model overfitting. This invention proposes a feature dimensionality reduction strategy based on single-cell-level biological mechanisms: scRNA-seq was used to analyze cellular heterogeneity in tea shoots, precisely identifying trichome cell subpopulations and extracting 63 core candidate genes specifically expressed that were highly correlated with the target phenotype (trichome density). During data quality control, signal-free variant sites with Mendelian errors or deletion rates exceeding one-third were strictly removed.

[0026] Within the trichome cell subset, one-way ANOVA was used to assess the impact of genotype variation on single-cell expression levels of target genes: for all valid SNPs located within a specific QTL interval, the aov(sc_expression~genotype_std) model was fitted, and its statistical significance P-value was calculated. To further compress feature dimensionality and eliminate collinearity caused by high linkage, this invention adopted a "top-1 per gene" strategy: in the single-cell association analysis of each candidate gene, only the single significant SNP with the smallest P-value (min_p_value) was extracted as the representative feature. Finally, 111 most significantly associated single-cell core regulatory SNP sites were successfully screened, constructing a Core-111 feature set with a clear mechanism explanation and extreme dimensionality reduction.

[0027] Alternatively, when assessing the association between genotype and expression levels, in addition to one-way ANOVA, linear mixed model (LMM) or MatrixeQTL algorithm can be used to correct for population structure and kinship. For the "Top-1 per gene" feature extraction strategy, to prevent randomness, it can be replaced by extracting tag SNPs within specific linkage disequilibrium (LD) blocks, or by relaxing the threshold and using a "Top-3" approach and a threshold truncation method based on false discovery rate (FDR) correction.

[0028] In this embodiment, the construction and evaluation of the ensemble learning prediction model are as follows: the extracted Core-111 feature set is used as the input matrix for machine learning. XTo capture the additive effects and nonlinear epistatic interactions of small-effect multi-gene pairs, an ensemble prediction model based on multi-algorithm fusion was constructed.

[0029] Parallel evaluation and validation: 5-fold cross-validation (5-foldCV) is introduced on the feature set to train and evaluate multiple algorithms in parallel, such as Ridge Regression, Lasso Regression, Random Forest, Gradient Boosting, and Support Vector Regression (SVR).

[0030] Ensemble algorithm optimization: By traversing all 2-combo and 3-combo combinations, the coefficient of determination (R²) is evaluated. 2 The results show that the model using the fusion of Ridge and Gradient Boosting (GB) performs best.

[0031] Integrated prediction logic: This optimal model employs an arithmetic mean method to fuse the outputs of linear and nonlinear base learners, effectively balancing the model's bias and variance. It is effective for any sample features. X Quantitative predicted value The calculation formula is as follows: in, This represents the predicted output of the ridge regression model. This represents the predicted output of the gradient boosting regression model.

[0032] Optionally, for the selection of base learners to capture linear / additive effects, in addition to Ridge Regression, Lasso Regression or ElasticNet can be equivalently replaced. For the selection of tree models to capture nonlinear interactions, in addition to Gradient Boosting Regression (GB), XGBoost, LightGBM, or RandomForest can be equivalently replaced. For the ensemble fusion logic, in addition to arithmetic averaging, weighted averaging (giving higher weights to better-performing base learners) or stacking / blending (training a meta-learner to combine the outputs of the primary models) can be used.

[0033] In this embodiment, the method for obtaining predicted values ​​of tea tree pubescence phenotypes after aligning the genotypes of the test samples in the feature space using the best ensemble learning model trained on the full dataset includes: The original dataset is input into the selected best ensemble learning model for final parameter iterative fitting. The original dataset contains all genotype matrices X and phenotype vectors Y for known phenotypes. The genotype data of the new sample to be tested is obtained and transposed. The feature columns of the new sample are strictly aligned to the column space dimension of the training set through a forced rearrangement algorithm. The aligned matrix is ​​input into the ensemble learning model, which is iteratively fitted with parameters. The model outputs the prediction results and generates a two-dimensional correlation report containing sample IDs and predicted phenotypic values.

[0034] The advantages of this invention are significant: 1. Possesses solid biological interpretability, resulting in a qualitative leap in overfitting prevention: Existing technologies often reduce dimensionality through pure mathematical transformations or "black box" importance scores (such as SHAP values) within deep learning, which easily lead to overfitting to meaningless background noise in small populations. This invention's feature extraction incorporates real biological mechanism verification (eQTL) of "whether the mutation causes changes in transcriptional expression," eliminating mutation sites without regulatory function. Under stringent conditions with only a few dozen small samples, its generalization stability far surpasses existing black box deep learning models.

[0035] 2. Capability to Capture Complex Genetic Effects Through a Two-Pronged Approach: Existing technologies using a single, heavy neural network (such as CNN) are prone to prediction bias when dealing with quantitative traits controlled by both additive and epistatic factors. This invention innovatively employs a parallel architecture of linear and nonlinear algorithms (Ridge+GB), complementing each other's advantages. Without increasing model complexity, it achieves a determination coefficient as high as 0.521, with accuracy superior to conventional models.

[0036] Example 2 The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The specific working process is as follows: (1) S1: Multi-omics data acquisition and feature initialization Single-cell transcriptome sequencing (scRNA-seq) was performed on the test materials 'Bai Mao 2' (hairy material) and 'Ku Wei Bai Mu Dan' (sparsely hairy material) to construct single-cell expression maps of tea shoots, and cell clustering and cell type annotation were completed. Based on this, the original expression count matrix and cell tag information of each cell group were extracted, and a cell type reference expression profile was constructed using InstaPrism software. Subsequently, this reference expression profile was combined with bulk RNA-seq data from 15 natural tea germplasm resources for deconvolution analysis to infer the composition ratio and expression contribution of each cell type in different bulk samples. That is, the overall expression of a certain gene in a bulk sample is considered as the superposition result of cell type-specific expression signals weighted by their cell abundance. in, Indicates the first In the bulk sample, the first The expression level of each gene, This indicates that the gene is in the first... Reference expression level in cells, Indicates the first In the nth sample The inferred proportion of cell types, This represents the error term. Using this method, the inferred cell type fraction (ICF) and the corresponding cell type-specific expression profile matrix for each bulk sample can be obtained.

[0037] After obtaining the ICF (Integrated Cell Formation Factor) for each cell group, correlation analysis was performed between them and the Trichotomy Index (TI). The results showed that among all cell groups, cluster13, i.e., the trichome cell group (TRI) in single-cell annotation, exhibited the most significant positive correlation with the trichome phenotype, with a correlation coefficient reaching [value missing]. The expression levels of TI and ICF13 were significantly higher than those of other cell groups. Based on the obtained TI and ICF13, further correlation screening was conducted on the gene expression levels of 15 tea germplasms, and 135 significantly related genes were finally identified. Among them, 63 genes were overlapping candidate genes that were significantly related to both TI and ICF13, which served as the core gene set for subsequent Core-111 screening.

[0038] Subsequently, transcriptome sequencing was performed on 'Gui Re 2', 'Jin Mu Dan', and their 77 F1 individuals to obtain population-scale RNA-seq data. Using the 'Shu Cha Zao' reference genome as a guide, cleanreads were first aligned using BWA, followed by format conversion, sorting, and index construction using SAMtools, and then labeled using GATK's MarkDuplicates tool. Based on this, variant detection was performed according to GATK best practices, including single-sample variant retrieval, gVCF merging, and joint genotyping, ultimately obtaining the original VCF file containing both parents and 77 F1 individuals. The original VCF file was then hard-filtered, and BCFtools was used to retain high-quality biallelic variant sites, forming the population genotypic data required for subsequent analysis.

[0039] After initial variant detection, Beagle software was used to fill and correct for missing genotypes in the population, with nthreads=8 to reduce incomplete genotypes caused by local low-depth sequencing or deletions. Based on this, a high-density genetic map was constructed using LEPMAP3. First, BCFtools and VCFtools were used to filter loci based on variant quality, alignment quality, sequencing depth, deletion rate, and minor allele frequency, eliminating low-quality, severely deleted, and rare allele loci, retaining only high-quality biallelic SNPs. Second, markers were classified into segregation types suitable for testcross strategies based on parental genotypes, such as lmxll, nnxnp, and hkxhk, and abnormal genotypes in offspring that clearly did not conform to parental inheritance patterns were uniformly marked as deletions. Subsequently, a chi-square test was used to eliminate significantly biased segregation sites to obtain a high-quality marker set suitable for linkage analysis. Finally, the SeparateChromosomes2 module of LEPMAP3 was used for linkage group division. By testing different LOD thresholds, as many effective markers as possible were retained while forming 15 linkage groups. Unassigned markers were then integrated using the JoinSingles2All module, and markers within each linkage group were ordered using the OrderMarkers2 module. A high-density genetic map containing 15 linkage groups and a total length of 1258.5 cM was finally constructed. After redundancy removal, 1,785 SNPs were retained as initial features, providing a high-quality genetic foundation for subsequent fine mapping and machine learning analysis. Figure 1 As shown.

[0040] Compared with existing technologies, this step does not directly rely on bulk RNA-seq differential expression analysis to screen candidate genes, nor does it directly use the massive SNPs of the whole genome for modeling. Instead, it first uses a single-cell reference map to analyze trichome-related cell components, then obtains the population-scale digital phenotype ICF13 through deconvolution, and combines phenotypic correlation information to identify core candidate genes. Finally, it conducts regulatory site mining based on a high-density genetic map, thereby improving the accuracy of cell type-specific signal analysis and the targeting of subsequent candidate site screening while retaining the advantage of population scale.

[0041] (2) S2: Core feature screening with single-cell precision (eQTL mapping) Traditional genome-wide prediction typically uses a large number of SNPs as model input. While this provides broad coverage, it easily introduces a large amount of redundant information unrelated to the target trait, leading to excessively high feature dimensionality, enhanced collinearity, and model overfitting. To address this, this invention proposes a feature dimensionality reduction strategy based on joint constraints of single-cell resolution information and population expression regulation information. Specifically, it first screens for core candidate genes for trichomes based on single-cell deconvolution and phenotypic correlation, then performs eQTL mapping around these core genes, and finally selects the optimal representative SNPs within significant regulatory intervals, thereby constructing high-information-density, low-redundancy machine learning input features.

[0042] Specifically, targeting the 63 core candidate genes obtained above, eQTL mapping analysis was conducted on their expression levels in the 'Gui Re 2' × 'Jin Mu Dan' F1 population. The expression level of each core gene in 77 F1 individuals was used as a quantitative trait, and SNP markers on a high-density genetic map were used as explanatory variables. A whole-genome scan was performed using the ICIM-ADD method in GACD v2 software, with a mapping step size of 1 cM. A LOD of 2.5 was used as a significance threshold to identify significant regulatory regions. After mapping, based on the physical distance between the eQTL site and the target gene, sites within a 1 Mb range upstream and downstream of the target gene were defined as cis-eQTLs, and the remaining sites were defined as trans-eQTLs. This method not only identifies regulatory regions significantly associated with the expression levels of core candidate genes but also further distinguishes between proximal and distal regulation, providing a basis for subsequent selection of key regulatory variants.

[0043] The key feature of this invention lies not in using eQTL analysis alone, but in the fact that the analysis targets are not all genes in the entire transcriptome, but 63 core genes obtained after multiple constraints including single-cell annotation, deconvolution analysis, and trichome phenotype correlation analysis. Therefore, the eQTL analysis target is more focused, and the obtained regulatory regions are more likely to be directly related to trichome development, avoiding the scattered and redundant results caused by unfocused scanning of the entire genome.

[0044] After obtaining significant eQTL intervals for 63 core candidate genes, all candidate SNPs were further extracted from the significant intervals corresponding to each target gene, and statistical screening was performed using genotype information from parents and the F1 population, target gene expression levels, and numerical phenotypic data. First, before candidate SNP screening, genotype-deficient sites in the population were filled using Beagle software based on the organized VCF file. Then, all SNP sites within the intervals were extracted based on the target genes and their corresponding localization intervals, and genotype information from parents and the F1 population, along with target expression level data, were integrated. For each site, the genetic error rate was calculated, which is the proportion of genotypes in offspring that do not conform to the parental inheritance pattern to the total number of valid observations at that site. in, This indicates the number of genotypes in offspring that do not conform to the parental inheritance pattern. This represents the total number of valid observations at that locus. When the genetic error rate at a locus exceeds 1 / 3, it is considered a low-quality locus and is removed. This step is used to exclude as many interfering loci as possible caused by misgenotyping, low sequencing quality, or abnormal recombination signals.

[0045] For the retained loci, a minimum of three valid samples are required, and at least two genotypes must be included to ensure that the locus meets the basic statistical comparison criteria. Subsequently, each candidate locus is genotype-standardized and coded. A one-way ANOVA is then performed, using the target gene expression level as the response variable and the locus genotype as the explanatory variable, to assess the explanatory power of different genotypes for differences in target gene expression levels. For a significant regulatory interval corresponding to the same target gene, the significance levels of all valid SNPs within that interval are compared, and only the single locus with the smallest p-value is retained as the representative SNP. In other words, this invention employs a "top-1" representative locus selection strategy per gene, rather than including all significant SNPs within an interval in the model. This significantly compresses feature dimensions, reduces redundancy caused by strongly linked loci, and retains the most representative regulatory loci strongly associated with the target expression level, thereby improving the compactness and stability of subsequent modeling features.

[0046] Through the above steps, 111 most significantly relevant core SNP loci were finally selected from the eQTL results of 63 core candidate genes, constructing a simplified Core-111 feature set. Furthermore, based on different levels of biological prior information, this invention also constructed another type of SNP feature set. Specifically, starting with the candidate genes corresponding to the three trichome-related modules M2, M9, and M11 identified by single-cell hdWGCNA, 1523 SNPs were obtained by screening their eQTL results, constructing the Module-1523 feature set, representing a wider range of candidate feature spaces. In contrast, Core-111, based on the 63 core candidate genes, further extracts key representative loci from their eQTL results, representing a simplified feature set constrained by multiple layers of evidence. The purpose of setting these two types of input features is to compare the modeling performance of wider-range candidate features and simplified core features in trichome trait prediction, thereby evaluating the impact of feature size and feature effectiveness on model performance.

[0047] Unlike existing technologies that directly use massive amounts of SNPs from the entire genome for prediction, this invention proposes a feature dimensionality reduction method with explicit biological prior constraints. Its core idea is not to input all SNPs into the model and let the algorithm learn on its own, but rather to further perform QTL analysis and screen core SNP sites through joint analysis of single-cell transcriptome data, conventional population transcriptome data, and phenotypic data. This method can significantly reduce feature dimensionality while maintaining biological relevance, improving the interpretability of candidate sites and the stability of subsequent prediction models. Figure 2 As shown, this step completes the screening of core regulatory sites based on single-cell precision information and constructs the Core-111 feature set for subsequent machine learning modeling.

[0048] (3) S3: Construction and evaluation of ensemble learning prediction models The Core-111 feature set extracted in step S2 is used as the input matrix for machine learning. ICF13 was used as the response variable. Six single models, as well as combinations of two and three models, were evaluated in parallel. The single models used included Ridge Regression, Lasso Regression, Random Forest Regression (RF), Gradient Boosting Regression (GB), Support Vector Regression (SVR), and K-Nearest Neighbors Regression (KNN). For RF and GB, the number of base learners was set to 100, and the random seed was set to 42. Cross-validation used a 5-fold setting: n_splits=5, shuffle=True, and random_state=42.

[0049] In each fold cross-validation, each individual model is first fitted using the training set, and the corresponding predicted value is obtained on the test set. Then, for combinations of two or three models, the arithmetic mean of the predictions from each base model is taken as the final predicted value for that combined model. For each individual model and model combination, the coefficient of determination is calculated on the test set. Root mean square error (RMSE) and prediction bias coefficient The results of the 50% discount were then compiled, and the Mean of each model was further statistically analyzed. Std ,Mean RMSE,Std RMSE,Mean and Std Then press Mean Sort from highest to lowest.

[0050] For model ensembles, the predicted values ​​are calculated using the average of the outputs of each base model. If a certain ensemble model consists of... If the model consists of several base models, then its predicted value for any sample is: in, Indicates the first Each base model for input features The predicted output. For the final optimal combination of the two models, Ridge+GB, its predicted value can be written as: Based on the 5-fold cross-validation results of the Core-111 and Module-1523 SNP feature sets, Ridge+GB is the best-performing model combination on the Core-111 dataset, with its Mean... The mean RMSE is 0.0072, which is 0.5185. The mean is 1.1168; Ridge+RF Mean of 0.5168 for Ridge+RF+GB The mean value is 0.4934, followed by Ridge+GB+KNN, Ridge+RF+KNN, and Ridge+Lasso+GB. On the Module-1523 dataset, Ridge+GB remains the optimal model combination, with a mean of 0.4934. The mean RMSE is 0.2518, and the mean RMSE is 0.0088. The value is 1.0859; Ridge+RF+GB and Ridge+RF ranked 2nd and 3rd respectively, Mean The mean values ​​are 0.2511 and 0.2152, respectively. Compared to Core-111, the mean values ​​of top-performing models on the Module-1523 dataset are significantly higher. Overall, it is relatively low, while the Mean RMSE is relatively high.

[0051] From the perspective of inter-fold fluctuations, the best model on the Core-111 dataset, Ridge+GB, has a Std value of... The Std RMSE is 0.1272, and the Std RMSE is 0.0019; the best model on the Module-1523 dataset, Ridge+GB, has the highest Std RMSE. The performance index (SMI) was 0.2149, and the Std RMSE was 0.0021. Considering all metrics, the Ridge+GB model on the Core-111 dataset was ultimately selected as the optimal model for subsequent phenotypic prediction and feature interpretation. Figure 3 , Figure 4 As shown.

[0052] (4) S4: Sample phenotype prediction and engineering output To apply the model to the early identification of new materials, this invention designs a feature space alignment mechanism. The first step is global parameter fitting. This involves using the complete original dataset (containing a matrix of all genotypes with known phenotypes). With phenotypic vectors The input is fed into the selected best model (Ridge+GB) for final training. Specifically, when the best model name contains "+", it is first split into the corresponding set of base model names, and then each base model is trained using fit(X, Y) on the complete original dataset. For the best combined model Ridge+GB, the Ridge model and the GB model are trained separately on the complete original dataset to obtain the final model for predicting new materials. This process can be represented as: in, and These represent the final Ridge and GB models trained on the complete original dataset, respectively.

[0053] Next, feature reindexing and alignment are performed. After obtaining the genotype data of the new samples to be tested, the data is read in using the group as the index column. Since the original new material genotype file has loci as rows and samples as columns, it is first transposed to ensure that the new sample matrix and the training matrix maintain the same orientation, i.e., samples as rows and SNP loci as columns. Subsequently, the training data matrix is ​​used... Using the column names as a standard, the transposed new sample matrix is ​​rearranged to ensure that the feature columns of the new samples are completely consistent with the column space of the training data. Redundant positions in the new samples that do not belong to the training column space are discarded during the rearrangement process. This process corresponds to `reindex(columns=X.columns)` in the code, and can be represented as: in, This represents the genotype matrix of the new sample after transposition. This represents the new sample input matrix after rearranging the columns of the training matrix.

[0054] Finally, the results are output. The aligned new sample matrix is ​​input into the optimal combination model (Ridge+GB), and the two sub-models, Ridge and GB, output predicted values ​​respectively. The two sets of predicted results are then averaged to obtain the final predicted phenotypic value of the new material. Subsequently, the new sample numbers and corresponding predicted values ​​were compiled into a two-dimensional results table for phenotypic prediction and ranking analysis of new materials, which can be directly used to assist molecular breeding decisions.

[0055] Compared with existing whole-genome prediction or traditional modular prediction technologies, this invention directly brings the following significant positive effects due to the aforementioned technical features: (1) Mechanism-oriented minimalist feature reduction effectively overcomes the curse of dimensionality: Existing techniques often directly use massive amounts of non-mechanistically screened variants, leading to overfitting. This invention introduces the eQTL expression regulation mechanism analysis mapped from tea pubescence-specific genes screened by single cells, accurately condensing the predicted input features to 111 core sites (Core-111) with clear biological regulatory significance. This significantly reduces the consumption of computational resources, fundamentally eliminates noise interference from irrelevant variants, and endows the prediction model with strong biological interpretability.

[0056] (2) Deep integration of multiple algorithms significantly improves the accuracy and stability of small sample prediction: Compared with the shortcomings of insufficient fitting ability of single algorithms, the ensemble model of this invention uses ridge regression and gradient boosting regression (GB) to form the optimal combination model (Ridge+GB), which achieves high accuracy and stability under small sample conditions. Reaching 0.521 It achieves a predictive efficacy of 0.780.

[0057] (3) Eliminate redundant dependencies and effectively prevent model overfitting: Compared with traditional modular prediction that blindly introduces a large number of variable features (such as the optimal combination model Ridge+GB when using the Module-1523 feature set), With a resolution of only 0.2518, this invention models using a compact core feature set (Core-111), reducing the model's dependence on redundant features and ensuring the capture of heritable variations.

[0058] Example 3 The present invention also provides a tea tree hair phenotype prediction system based on machine learning. The system is used to implement the method described in Embodiment 1. The system includes: a data acquisition module, a site screening module, a model building module, and a phenotype prediction module. The data acquisition module is used to perform single-cell transcriptome sequencing (scRNA-seq) on the test material to obtain cell type-specific genes; and to perform transcriptome sequencing on the test population to construct a high-density genetic map and extract a set of single nucleotide polymorphism (SNP) markers. The site screening module is used to perform eQTL mapping of expression quantitative trait sites for candidate genes specifically expressed by trichome cells selected by single-cell sequencing and trichome phenotype, and to identify and extract associated SNP sites. The model building module is used to construct an ensemble learning model based on the fusion of multiple algorithms, based on associated SNP sites. The phenotypic prediction module is used to obtain the predicted values ​​of tea tree hair phenotype by aligning the genotypes of the test samples in the feature space using the best ensemble learning model trained on the full dataset.

[0059] In this embodiment, the site screening module includes: a single-cell clustering unit and a feature dimensionality reduction unit; The single-cell clustering unit is used to analyze the heterogeneity of tea tree shoot cells using scRNA-seq, identify trichome cell subpopulations, and extract several specific expression core candidate genes related to the target phenotype. The feature reduction unit is used to combine high-density genetic maps and perform eQTL mapping on trichome-specific expression genes screened at the single-cell level. It screens out 111 core single nucleotide polymorphism (SNP) sites that are most significantly associated with trichome development from the whole genome and constructs the dimensionality-reduced Core-111 feature set.

[0060] In this embodiment, combining a high-density genetic map, the expression quantitative trait loci (eQTLs) of trichome-specific expression genes screened at the single-cell level are mapped to the whole genome to screen 111 core single nucleotide polymorphism (SNP) sites most significantly associated with trichome development. The process of constructing the dimensionality-reduced Core-111 feature set includes: During the data quality control phase, non-signal variant sites with Mendelian errors or a deletion rate exceeding 1 / 3 are removed. Within the trichome cell subset, the effect of genotypic variation on the single-cell expression level of target genes was assessed using one-way ANOVA. For a significant regulatory region corresponding to the same target gene, the significance levels of all SNPs within that region are compared. Only the single site with the smallest P-value is retained as the representative SNP. Finally, 111 single-cell core regulatory SNP sites with the most significant correlation are selected, and a Core-111 feature set with mechanistic explanation and extreme dimensionality reduction is constructed.

[0061] In this embodiment, the ensemble learning model achieves quantitative prediction of the tea plant hair phenotype by weighted ensemble of base learners, including ridge regression and gradient boosting regression. Specifically: Quantitative prediction value for any sample feature X The calculation formula is: ; in, This represents the predicted output of the ridge regression model. This represents the predicted output of the gradient boosting regression model.

[0062] In this embodiment, the phenotype prediction module includes: a global parameter fitting unit, a feature rearrangement and alignment unit, and a result output unit; The global parameter fitting unit is used to input the original dataset into the selected best ensemble learning model for final training. The original dataset contains all genotype matrices X and phenotype vectors Y for known phenotypes. Specifically: when the best model name contains "+", it is first split into the corresponding set of base model names, and then each base model is trained using fit(X, Y) on the complete original dataset. For the best combined model Ridge+GB, the Ridge model and the GB model are trained on the complete original dataset respectively to obtain the final model for predicting new materials, represented as: ; in, and These represent the final Ridge and GB models trained on the complete original dataset, respectively. The feature rearrangement and alignment unit is used to acquire and transpose the genotype data of the new sample to be tested. Through a forced rearrangement algorithm, the feature columns of the new sample are strictly aligned to the column space dimension of the training set. Specifically: after acquiring the genotype data of the new sample to be tested, the data is read in using the group as the index column; it is first transposed to ensure that the new sample matrix and the training matrix maintain the same orientation, i.e., samples as rows and SNP sites as columns; subsequently, the training data matrix... Using the column names as criteria, the transposed new sample matrix is ​​rearranged; any redundant sites in the new sample that do not belong to the training column space are discarded during the rearrangement process; this process is represented as: ; in, This represents the genotype matrix of the new sample after transposition. This represents the new sample input matrix after rearranging according to the column order of the training matrix. Indicates a column; The results output unit is used to input the aligned matrix into the ensemble learning model after iterative fitting of the parameters, output the prediction results, and generate a two-dimensional correlation report containing sample IDs and predicted phenotypic values. Specifically: the aligned new sample matrix is ​​input into the best combination model Ridge+GB, and the two sub-models Ridge and GB output the predicted values ​​respectively. The two sets of prediction results are then averaged to obtain the final predicted phenotypic value of the new material. ; Subsequently, the new sample numbers and corresponding predicted values ​​were compiled into a two-dimensional results table for phenotypic prediction and ranking analysis of new materials, which can be directly used to assist molecular breeding decisions.

[0063] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting the pubescent phenotype of tea trees based on machine learning, characterized in that, The method includes: Single-cell transcriptome sequencing (scRNA-seq) was performed on the test material to obtain cell type-specific genes; transcriptome sequencing was performed on the test population to construct a high-density genetic map and extract a set of single nucleotide polymorphism (SNP) markers. For candidate genes specifically expressed by trichome cells identified through single-cell sequencing and trichome phenotype screening, eQTL mapping of expression quantitative trait sites was performed to identify and extract associated SNP sites. Based on associated SNP sites, an ensemble learning model based on the fusion of multiple algorithms is constructed. By using the best ensemble learning model trained on the full dataset, the genotypes of the test samples are aligned in feature space to obtain the predicted values ​​of the tea tree hair phenotype.

2. The method according to claim 1, characterized in that, For candidate genes specifically expressed in trichome cells identified through single-cell sequencing and trichome phenotype screening, methods for performing eQTL mapping to identify and extract associated SNP sites include: scRNA-seq was used to analyze the cellular heterogeneity of tea tree shoots, identify trichome cell subpopulations, and extract several core candidate genes specifically expressed in relation to the target phenotype. By combining high-density genetic maps, eQTL mapping of expression quantitative trait loci was performed on trichome-specific expression genes screened at the single-cell level. 111 core single nucleotide polymorphism (SNP) loci that are most significantly associated with trichome development were screened from the whole genome, and a dimensionality-reduced Core-111 feature set was constructed.

3. The method according to claim 2, characterized in that, Combining high-density genetic maps, eQTL mapping was performed on trichome-specific expression genes screened at the single-cell level. This allowed for the screening of 111 core single nucleotide polymorphisms (SNPs) most significantly associated with trichome development from the whole genome. The method for constructing the dimensionality-reduced Core-111 feature set includes: During the data quality control phase, non-signal variant sites with Mendelian errors or a deletion rate exceeding 1 / 3 are removed. Within the trichome cell subset, the effect of genotypic variation on the single-cell expression level of target genes was assessed using one-way ANOVA. For a significant regulatory region corresponding to the same target gene, the significance levels of all SNPs within that region are compared. Only the single site with the smallest P-value is retained as the representative SNP. Finally, 111 single-cell core regulatory SNP sites with the most significant correlation are selected, and a Core-111 feature set with mechanistic explanation and extreme dimensionality reduction is constructed.

4. The method according to claim 1, characterized in that, The ensemble learning model achieves quantitative prediction of tea plant hair phenotype by weighted ensemble of base learners, including ridge regression and gradient boosting regression. Specifically: Quantitative prediction value for any sample feature X The calculation formula is: ； in, This represents the predicted output of the ridge regression model. This represents the predicted output of the gradient boosting regression model.

5. The method according to claim 1, characterized in that, Methods for obtaining predicted values ​​of tea tree pubescence phenotypes by using the best ensemble learning model trained on the full dataset and aligning the genotypes of the test samples in the feature space include: The original dataset is input into the selected optimal ensemble learning model for final training. The original dataset contains the genotype matrix X and phenotype vector Y for all known phenotypes. Specifically: when the optimal model name contains "+", it is first split into the corresponding set of base model names, and then each base model is trained using fit(X,Y) on the complete original dataset. For the optimal combined model Ridge+GB, the Ridge model and the GB model are trained on the complete original dataset respectively to obtain the final model for predicting new materials, represented as: ； in, and These represent the final Ridge and GB models trained on the complete original dataset, respectively. The genotype data of the new samples to be tested are obtained and transposed. A forced rearrangement algorithm is used to strictly align the feature columns of the new samples to the column space dimension of the training set. Specifically: after obtaining the genotype data of the new samples, the data is read in using the group as the index column; it is first transposed to ensure that the new sample matrix and the training matrix maintain the same orientation, i.e., samples as rows and SNP sites as columns; subsequently, the training data matrix... Using the column names as criteria, the transposed new sample matrix is ​​rearranged; any redundant sites in the new sample that do not belong to the training column space are discarded during the rearrangement process; this process is represented as: ； in, This represents the genotype matrix of the new sample after transposition. This represents the new sample input matrix after rearranging according to the column order of the training matrix. Indicates a column; The aligned matrix is ​​input into the ensemble learning model after iterative fitting of parameters, and the output prediction results are generated to produce a two-dimensional correlation report containing sample IDs and predicted phenotypic values. Specifically: the aligned new sample matrix is ​​input into the optimal combination model Ridge+GB, and the two sub-models Ridge and GB output prediction values ​​respectively. The two sets of prediction results are then averaged to obtain the final predicted phenotypic value of the new material. ； Subsequently, the new sample numbers and corresponding predicted values ​​were compiled into a two-dimensional results table for phenotypic prediction and ranking analysis of new materials, which can be directly used to assist molecular breeding decisions.

6. A machine learning-based tea tree pubescence phenotypic prediction system, said system being used to implement the method described in any one of claims 1-5, characterized in that, The system includes: a data acquisition module, a site selection module, a model building module, and a phenotypic prediction module; The data acquisition module is used to perform single-cell transcriptome sequencing (scRNA-seq) on the test material to obtain cell type-specific genes; and to perform transcriptome sequencing on the test population to construct a high-density genetic map and extract a set of single nucleotide polymorphism (SNP) markers. The site screening module is used to perform eQTL mapping on candidate genes specifically expressed by trichome cells selected by single-cell sequencing and trichome phenotype, and to identify and extract associated SNP sites. The model building module is used to build an ensemble learning model based on the fusion of multiple algorithms based on associated SNP sites. The phenotypic prediction module is used to obtain tea tree hair phenotypic prediction values ​​by aligning the genotypes of the test samples in the feature space using the best ensemble learning model trained on the full dataset.

7. The system according to claim 6, characterized in that, The site screening module includes: a single-cell clustering unit and a feature dimensionality reduction unit; The single-cell clustering unit is used to analyze the heterogeneity of tea tree shoot cells using scRNA-seq, identify trichome cell subpopulations, and extract several specific expression core candidate genes related to the target phenotype. The feature reduction unit is used to combine a high-density genetic map, perform eQTL mapping on trichome-specific expression genes screened at the single-cell level, screen out 111 core single nucleotide polymorphism (SNP) sites most significantly associated with trichome development from the whole genome, and construct a dimensionality-reduced Core-111 feature set.

8. The system according to claim 7, characterized in that, Combining high-density genetic maps, the process of mapping quantitative trait loci (eQTLs) of trichome-specific expression genes screened at the single-cell level, and identifying 111 core single nucleotide polymorphism (SNP) sites most significantly associated with trichome development from the whole genome, and constructing the dimensionality-reduced Core-111 feature set includes: During the data quality control phase, non-signal variant sites with Mendelian errors or a deletion rate exceeding 1 / 3 are removed. Within the trichome cell subset, the effect of genotypic variation on the single-cell expression level of target genes was assessed using one-way ANOVA. For a significant regulatory region corresponding to the same target gene, the significance levels of all SNPs within that region are compared. Only the single site with the smallest P-value is retained as the representative SNP. Finally, 111 single-cell core regulatory SNP sites with the most significant correlation are selected, and a Core-111 feature set with mechanistic explanation and extreme dimensionality reduction is constructed.

9. The system according to claim 6, characterized in that, The ensemble learning model achieves quantitative prediction of tea plant hair phenotype by weighted ensemble of base learners, including ridge regression and gradient boosting regression. Specifically: Quantitative prediction value for any sample feature X The calculation formula is: ； in, This represents the predicted output of the ridge regression model. This represents the predicted output of the gradient boosting regression model.

10. The system according to claim 6, characterized in that, The phenotypic prediction module includes: a global parameter fitting unit, a feature rearrangement and alignment unit, and a result output unit; The global parameter fitting unit is used to input the original dataset into the selected best ensemble learning model for final training. The original dataset contains all genotype matrices X and phenotype vectors Y for known phenotypes. Specifically: when the best model name contains "+", it is first split into the corresponding set of base model names, and then each base model is trained using fit(X, Y) on the complete original dataset. For the best combined model Ridge+GB, the Ridge model and the GB model are trained on the complete original dataset respectively to obtain the final model for predicting new materials, expressed as: ； in, and These represent the final Ridge and GB models trained on the complete original dataset, respectively. The feature rearrangement and alignment unit is used to acquire and transpose the genotype data of the new sample to be tested. Through a forced rearrangement algorithm, the feature columns of the new sample are strictly aligned to the column space dimension of the training set. Specifically: after acquiring the genotype data of the new sample to be tested, the data is read in using the group as the index column; it is first transposed to ensure that the new sample matrix and the training matrix maintain the same orientation, i.e., samples as rows and SNP sites as columns; subsequently, the training data matrix... Using the column names as criteria, the transposed new sample matrix is ​​rearranged; any redundant sites in the new sample that do not belong to the training column space are discarded during the rearrangement process; this process is represented as: ； in, This represents the genotype matrix of the new sample after transposition. This represents the new sample input matrix after rearranging according to the column order of the training matrix. Indicates a column; The result output unit is used to input the aligned matrix into the ensemble learning model after iterative fitting of the parameters, output the prediction results, and generate a two-dimensional correlation report containing sample IDs and predicted phenotypic values. Specifically: the aligned new sample matrix is ​​input into the optimal combination model Ridge+GB, and the two sub-models Ridge and GB output the predicted values ​​respectively. The two sets of prediction results are then averaged to obtain the final predicted phenotypic value of the new material. ； Subsequently, the new sample numbers and corresponding predicted values ​​were compiled into a two-dimensional results table for phenotypic prediction and ranking analysis of new materials, which can be directly used to assist molecular breeding decisions.

Images