Target character prediction method suitable for low-depth sequencing and whole genome selection model suitable for low-depth sequencing

Abstract

The invention provides a target character prediction method suitable for low-depth sequencing and a whole genome selection model suitable for low-depth sequencing, and relates to the technical field of biology. According to the target character prediction method suitable for low-depth sequencing and the whole-genome selection model provided by the invention, the target character of the to-be-detected sample can be accurately predicted under a limited data volume by combining low-depth whole-genome re-sequencing data with a machine learning technology, and the genome selection efficiency can be improved by matching the two; the time and economic cost of breeding or medical research are reduced, and the method is particularly suitable for large-scale genetic analysis scenes needing rapid iteration.

Description

Technical Field

[0001] This invention relates to the field of biotechnology, and in particular to a method for predicting a target trait suitable for low-depth sequencing and a genome-wide selection model suitable for low-depth sequencing. Background Technology

[0002] Genomic selection (GS) is a technique that uses statistical models to predict individual breeding values ​​by analyzing the association between high-density genetic markers (such as SNPs) and phenotypic data across the entire genome, thus achieving early and precise breeding. However, it has the following drawbacks:

[0003] (1) Genotyping methods are costly or have a limited number of markers.

[0004] Currently, the genotyping methods used for whole-genome selection are usually whole-genome resequencing and SNP microarrays. The former is expensive and not suitable for large-scale sample and population studies of large genome species, while the latter obtains tens to hundreds of SNP loci.

[0005] (2) The dilemma of high-dimensional data processing

[0006] Traditional linear models (such as GBLUP) rely on matrix operations. When faced with millions of SNP tags, the computational efficiency of traditional models decreases due to the complexity of matrix inversion, and the computation time increases exponentially.

[0007] (3) Does not support nonlinear relationship modeling

[0008] Traditional models can only fit the linear relationship between markers and phenotypes, and cannot capture epistasis or environment-gene interactions, which limits the prediction accuracy of complex traits (such as disease resistance and yield).

[0009] Therefore, the key to genetic evaluation is how to obtain high-density genotype information from large-scale populations at low cost and how to construct GS models for efficient genetic assessment.

[0010] In view of this, the present invention is hereby proposed. Summary of the Invention

[0011] The primary objective of this invention is to provide a method for predicting target traits applicable to low-depth sequencing, thereby addressing the aforementioned technical problems.

[0012] A second objective of this invention is to provide a genome-wide selection model suitable for low-depth sequencing.

[0013] To achieve the above objectives, the following technical solution is adopted:

[0014] In a first aspect, the present invention provides a method for predicting a target trait suitable for low-depth sequencing, comprising the following steps:

[0015] a. Obtain phenotypic data and whole-genome sequencing data of the samples;

[0016] b. Genotype filling is performed on the whole genome sequencing data to obtain genotype data, and SNP loci associated with the target trait are screened from the genotype data;

[0017] c. Using the SNP sites as features, train a classification model using the phenotypic data and genotypic data to obtain a trained classification model;

[0018] d. Genotyping is performed on the whole-genome sequencing data of the sample to be tested, and then the trained classification model is used for classification to predict the target trait of the sample to be tested;

[0019] The sequencing depth of the whole genome sequencing data was 0.5×-2×.

[0020] As a further technical solution, in step b, a reference panel is constructed using GLIMPSE2 (2.0.0) software for high-depth sequencing data (>15X), and then the whole genome sequencing data is filled with genotypes based on the constructed reference panel.

[0021] As a further technical solution, during the genotype filling process, a hidden Markov model algorithm is used to infer the genotype of each sample at each site.

[0022] As a further technical solution, in step b, a genome-wide association analysis method is used to screen out SNP sites associated with the target trait.

[0023] As a further technical solution, the classification model includes LightGBM, random forest, deep neural network, support vector machine or Bayesian model.

[0024] Secondly, the present invention provides a whole-genome selection model suitable for low-depth sequencing, including an acquisition module, a preprocessing module, and a classification module;

[0025] The acquisition module is used to acquire whole genome sequencing data of the sample to be tested;

[0026] The preprocessing module is used to fill the genotypes in the whole genome sequencing data of the sample to be tested;

[0027] The classification module is used to input the data processed by the preprocessing module into the pre-trained classification model, and perform classification processing through the classification model to predict the target trait of the sample to be tested;

[0028] The classification model was trained using the following method:

[0029] a. Obtain phenotypic data and whole-genome sequencing data of the samples;

[0030] b. Genotype filling is performed on the whole genome sequencing data to obtain genotype data, and SNP loci associated with the target trait are screened from the genotype data;

[0031] c. Using the SNP sites as features, train a classification model using the phenotypic data and genotypic data to obtain a trained classification model;

[0032] The sequencing depth of the whole genome sequencing data was 0.5×-2×.

[0033] As a further technical solution, in step b, a reference panel is constructed using GLIMPSE2 (2.0.0) software for high-depth sequencing data (>15X), and then the whole genome sequencing data is filled with genotypes based on the constructed reference panel.

[0034] As a further technical solution, during the genotype filling process, a hidden Markov model algorithm is used to infer the genotype of each sample at each site.

[0035] As a further technical solution, in step b, a genome-wide association analysis method is used to screen out SNP sites associated with the target trait.

[0036] As a further technical solution, the classification model includes LightGBM, random forest, deep neural network, support vector machine or Bayesian model.

[0037] Compared with existing technologies, the prediction method and genome-wide selection model for target traits suitable for low-depth sequencing provided by this invention have the following advantages:

[0038] (1) Advantages of genotyping method: The reference group is constructed by filling the data after low-depth whole genome resequencing. Under the same sequencing cost, a larger sample size can be included in the reference group, which effectively improves the prediction accuracy of whole genome selection.

[0039] (2) Advantages of genomic selection models: Machine learning models can efficiently process high-dimensional sparse data generated by low-depth sequencing, and can optimize feature selection and model training by using their gradient boosting mechanism, achieving accurate predictions even with limited data. The combination of the two can significantly improve the efficiency of genomic selection, reduce the time and economic costs of breeding or medical research, and is especially suitable for large-scale genetic analysis scenarios that require rapid iteration. Attached Figure Description

[0040] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0041] Figure 1 The filling accuracy results for different MAF sites (R) 2 );

[0042] Figure 2 This is a flowchart of the method for predicting target traits applicable to low-depth sequencing according to the present invention;

[0043] Figure 3 Test results for different genome-wide selection models (R) 2 ). Detailed Implementation

[0044] The embodiments and examples of the present invention will be described in detail below. However, those skilled in the art will understand that the following embodiments and examples are for illustrative purposes only and should not be considered as limiting the scope of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Unless otherwise specified, conventional conditions or conditions recommended by the manufacturer shall apply. Reagents or instruments whose manufacturers are not specified are all commercially available conventional products.

[0045] In a first aspect, the present invention provides a method for predicting a target trait suitable for low-depth sequencing, comprising the following steps:

[0046] a. Obtain phenotypic data and whole-genome sequencing data of the samples;

[0047] b. Genotype filling is performed on the whole genome sequencing data to obtain genotype data, and SNP loci associated with the target trait are screened from the genotype data;

[0048] c. Using the SNP sites as features, train a classification model using the phenotypic data and genotypic data to obtain a trained classification model;

[0049] d. Genotyping is performed on the whole-genome sequencing data of the sample to be tested, and then the trained classification model is used for classification to predict the target trait of the sample to be tested;

[0050] The sequencing depth of the whole genome sequencing data was 0.5×-2×.

[0051] Low-coverage whole-genome sequencing (LcWGS) is a low-cost genomic analysis method based on high-throughput sequencing technology. This invention utilizes low-depth whole-genome sequencing on a large number of samples, leveraging linkage disequilibrium principles and combining a high-quality reference panel to perform genotype imputation on the sequencing data, thereby obtaining high-density single nucleotide polymorphism (SNP) markers across the entire genome. The method can obtain millions of SNP markers per sample. Subsequently, genome-wide association analysis (GWAS) can be used to screen for markers significantly associated with the target trait. A classification model (such as LightGBM) is then used for model training, and the trained model is used to predict the target trait of the sample. The low cost of low-coverage whole-genome sequencing makes it suitable for large-scale breeding populations. Combined with the efficient computational power of machine learning, it can significantly reduce time and economic costs, especially suitable for large-scale genetic analysis scenarios requiring rapid iteration. It solves the problems of traditional SNP chips obtaining limited loci and traditional GS models being limited by algorithmic complexity and unable to handle large-scale molecular markers.

[0052] In this invention, the sequencing depth of the whole genome sequencing data is preferably 1×.

[0053] The inventors have discovered that 1× sequencing has the following advantages:

[0054] a: Significantly reduces sequencing costs

[0055] Cost advantage: 1× sequencing requires only 1 / 5 the sequencing depth of traditional 5× sequencing, meaning that sequencing costs can be significantly reduced in large-scale sample analysis. This cost advantage is particularly important for research projects that require processing large numbers of samples, significantly improving the economics and scalability of the research.

[0056] b: Stable 1× coverage

[0057] Based on preliminary experiments with hundreds of samples, the library construction method of enzyme digestion was determined to achieve a stable high coverage of 1×, with the coverage remaining stable at around 58%.

[0058] c: Filling technology achieves results comparable to 15×.

[0059] Through efficient fill-in algorithms, we can effectively fill in blank areas in the data, building upon 1× sequencing, thereby improving data integrity and accuracy. This allows 1× sequencing to obtain SNP markers covering the entire genome, achieving a level comparable to 15× sequencing. Figure 1As shown: Low-depth whole-genome sequencing (LcWGS, ~1X) was performed on 100 pig samples. A reference panel was constructed using GLIMPSE2 (2.0.0) software (data sources: authoritative studies, public databases, and self-tested samples, totaling 3108 pigs, with an average sequencing depth of 20X). Genotyping was then performed on the whole-genome sequencing data based on the constructed reference panel. The filled genotypes were compared with the corresponding high-depth (WGS, ~15X) genotype data. The results showed that the accuracy of loci with a maf (minor allele frequency: the frequency of a less common allele at a locus in population genetics) greater than 0.05 was over 90%, and the accuracy of loci with a maf greater than 0.1 was over 95%.

[0060] In this invention, the phenotypic data includes target trait data.

[0061] In some optional implementations, the acquisition of phenotypic data includes data acquisition and data correction; wherein, data acquisition may, for example, utilize specific equipment or technology to systematically measure and record the morphological, physiological or behavioral characteristics of the target organism to ensure the comprehensiveness and accuracy of the data; data correction is to correct the acquired data, for example, by standardization processing and calibration algorithms, to denoise, correct errors and standardize the acquired phenotypic data to improve the reliability and consistency of the data.

[0062] In some alternative implementations, the whole-genome sequencing data is the raw data after filtering to remove low-quality bases, adapter sequences, and contaminating sequences.

[0063] In some alternative implementations, in step b, a reference panel is constructed using GLIMPSE2 (2.0.0) software for the high-depth sequencing data (>15X), and then the whole genome sequencing data is genotype-filled based on the constructed reference panel.

[0064] In some optional implementations, during the genotype filling process, a Hidden Markov Model (HMM) algorithm is used to infer the genotype of each sample at each locus to obtain high-precision genotype data.

[0065] In some alternative implementations, step b involves using genome-wide association analysis to screen for SNP loci associated with the target trait.

[0066] In some alternative implementations, the classification model includes LightGBM, random forest, deep neural network, support vector machine, or Bayesian model, with LightGBM being preferred. The inventors have found that using LightGBM as the classification model yields more accurate predictions.

[0067] In some optional embodiments, the method for predicting target traits suitable for low-depth sequencing provided by the present invention has the following process: Figure 2 As shown, it includes the following steps:

[0068] (1) Phenotypic collection: systematically measure and record the morphological, physiological or behavioral characteristics of the target organism (reference population) using specific equipment or technology to ensure the comprehensiveness and accuracy of the data.

[0069] (2) Phenotypic data correction: Through standardization processing and calibration algorithms, the collected phenotypic data is denoised, error corrected and standardized to improve the reliability and consistency of the data.

[0070] (3) Genotyping: The constructed library is then subjected to high-throughput sequencing. The sequencing platform can be the DNBSEQ T7 platform, set to PE150 mode, and the sequencing depth is 1X.

[0071] (4) Data filtering and quality control: The raw image data obtained from sequencing is converted into sequence data through base identification. The raw data is then filtered to remove low-quality bases, adapter sequences, and contaminating sequences to obtain high-quality, clean data. The filtering methods are as follows: a) Sequencing fragments containing adapter sequences need to be filtered out; b) When the content of N in a single-end sequencing fragment exceeds 10% of the length of the sequencing fragment, the sequencing fragment needs to be removed; c) When the number of low-quality (≤5) bases in a single-end sequencing fragment exceeds 50% of the length of the sequencing fragment, the sequencing fragment needs to be removed.

[0072] (5) Genotype filling: Using GLIMPSE2 (2.0.0) software, genotype filling was performed on low-depth sequencing data using a reference panel constructed from high-depth sequencing data. The genotype of each sample at each locus was inferred using a hidden Markov model algorithm to obtain high-precision genotype data.

[0073] (6) SNP marker screening: Select SNP markers related to the target trait from the filled genotype data. Genome-wide association analysis can be used to screen for significantly associated SNP sites.

[0074] (7) Model building: Set the basic parameters of the LightGBM model, such as the objective function (e.g., regression or classification), learning rate, number of trees, number of leaf nodes, etc.

[0075] (8) Model training: Divide the preprocessed data into training set, validation set and test set in a ratio of 6:2:2. Use the training set data to train the LightGBM model. Iteratively optimize the model parameters to enable the model to accurately predict the target trait.

[0076] (9) Model evaluation: Use test set data to evaluate the trained model. Commonly used evaluation metrics include root mean square error (RMSE), coefficient of determination (R²), correlation coefficient, etc.

[0077] (10) Model optimization: The hyperparameters of the LightGBM model are tuned using methods such as grid search, random search or Bayesian optimization to improve the model performance.

[0078] Secondly, the present invention provides a whole-genome selection model suitable for low-depth sequencing, including an acquisition module, a preprocessing module, and a classification module;

[0079] The acquisition module is used to acquire whole genome sequencing data of the sample to be tested;

[0080] The preprocessing module is used to fill the genotypes in the whole genome sequencing data of the sample to be tested;

[0081] The classification module is used to input the data processed by the preprocessing module into the pre-trained classification model, and perform classification processing through the classification model to predict the target trait of the sample to be tested;

[0082] The classification model was trained using the following method:

[0083] a. Obtain phenotypic data and whole-genome sequencing data of the samples;

[0084] b. Genotype filling is performed on the whole genome sequencing data to obtain genotype data, and SNP loci associated with the target trait are screened from the genotype data;

[0085] c. Using the SNP sites as features, train a classification model using the phenotypic data and genotypic data to obtain a trained classification model;

[0086] The sequencing depth of the whole genome sequencing data was 0.5×-2×.

[0087] The whole-genome selection model provided by this invention, which is suitable for low-depth sequencing, utilizes low-depth whole-genome resequencing data combined with machine learning technology to achieve accurate prediction of the target trait of the sample under limited data volume, improves the efficiency of genome selection, and reduces the time and economic costs of breeding or medical research. It is especially suitable for large-scale genetic analysis scenarios that require rapid iteration.

[0088] In some alternative implementations, in step b, a reference panel is constructed using GLIMPSE2 (2.0.0) software for the high-depth sequencing data (>15X), and then the whole genome sequencing data is genotype-filled based on the constructed reference panel.

[0089] In some optional implementations, during the genotype filling process, a hidden Markov model algorithm is used to infer the genotype of each sample at each locus to obtain high-precision genotype data.

[0090] In some alternative implementations, step b involves using genome-wide association analysis to screen for SNP loci associated with the target trait.

[0091] In some optional implementations, the classification model includes LightGBM, random forest, deep neural network, support vector machine, or Bayesian model, with LightGBM being preferred. LightGBM can efficiently process high-dimensional sparse data generated by low-depth sequencing, and its gradient boosting mechanism optimizes feature selection and model training, achieving accurate predictions even with limited data.

[0092] The present invention will be further illustrated below with specific embodiments. However, it should be understood that these embodiments are merely for the purpose of more detailed illustration and should not be construed as limiting the present invention in any way.

[0093] Example 1:

[0094] A method for predicting reproductive traits in Large White pigs, using 2000 Large White pigs from a breeding company as the experimental population, includes the following steps:

[0095] (1) Phenotypic data collection: The reproductive traits of 2,000 Large White pigs were systematically measured and recorded. The results are shown in Table 1.

[0096] Table 1 Reproductive traits of 2000 Large White pigs

[0097] Total number of children Number of samples 0-5 243 5-10 403 10~15 1024 15-20 330

[0098] (2) Genotyping: The sample is subjected to genome sequencing. The sequencing platform can be selected as DNBSEQ T7 platform, set to PE150 mode, and the sequencing depth is 1X.

[0099] (3) Data filtering and quality control: The raw image data obtained from sequencing is converted into sequence data through base identification. The raw data is then filtered to remove low-quality bases, adapter sequences, and contaminating sequences to obtain high-quality, clean data. The filtering methods are as follows: a) Sequencing fragments containing adapter sequences need to be filtered out; b) When the content of N in a single-end sequencing read exceeds 10% of the length of the sequencing fragment, the sequencing fragment needs to be removed; c) When the number of low-quality (≤5) bases in a single-end sequencing fragment exceeds 50% of the length of the sequencing fragment, the sequencing fragment needs to be removed.

[0100] (4) Genome alignment: The low-depth resequencing data was aligned to the reference genome (e.g., pig version 11.1) using BWA software. Genotype filling was then performed using GLIMPSE2 (2.0.0) software. A reference panel was constructed using high-depth sequencing data (data from 3108 pigs from authoritative studies, public databases, and self-tested samples, with an average sequencing depth of 20X). The filtering conditions were as follows:

[0101] ① Eliminate multiple alleles;

[0102] ② Remove sites with MAF < 0.01;

[0103] ③ Retain sites with a deletion rate ≥ 0.9.

[0104] By using the Hidden Markov Model (HMM) algorithm, the genotype of each sample at each locus is inferred, resulting in high-precision genotype data.

[0105] (5) SNP marker screening: The processed data were divided into training set, validation set and test set in a 6:2:2 ratio. From the training set, genome-wide association analysis (GWAS) was used to screen out significantly associated SNP sites.

[0106] (6) Model construction: Set the basic parameters of the LightGBM model: 'objective': 'binary', 'boosting_type': 'gbdt', 'metric': 'auc', 'learning_rate': 0.05, 'n_estimators': 1000.

[0107] (7) Model training: Use the training set data to train the LightGBM model and obtain a trained classification model.

[0108] (8) Model validation: The trained model is evaluated using validation set data. Root mean square error (RMSE) and coefficient of determination (R²) are used as evaluation metrics. Model parameters are continuously adjusted based on the validation results to achieve optimal results. The results are as follows:

[0109] Model validation results (validation set):

[0110] - Root mean square error (RMSE): 1.2;

[0111] - Coefficient of determination (R²): 0.55;

[0112] This indicates that the model has good stability and generalization ability.

[0113] (9) Model testing: The model was tested using the test set, and the results are as follows:

[0114] Accuracy: R² = 0.52; Time required: 1 day.

[0115] Comparative Example 1

[0116] The training set samples in Example 1 were analyzed using the traditional GBLUP model. Then, the sample traits of the test set data in Example 1 were predicted, and the results are as follows:

[0117] Accuracy: R² = 0.2; Time required: 3 days.

[0118] The results of Example 1 and Comparative Example 1 show that the method provided by this invention can significantly improve the prediction accuracy of Large White pigs at a depth of 1× (R²=0.52), while greatly shortening the model training time (comparing 1 day in Example 1 with 3 days in Comparative Example 1). Therefore, this method can effectively utilize low-depth sequencing data for molecular breeding and trait prediction research.

[0119] Example 2

[0120] The method for predicting reproductive traits in Large White pigs differs from Example 1 in that LightGBM is replaced with rrBLUP.

[0121] Example 3

[0122] The method for predicting reproductive traits in Large White pigs differs from that in Example 1 in that LightGBM is replaced with Support Vector Machine (SVR).

[0123] Example 4

[0124] The method for predicting reproductive traits in Large White pigs differs from that in Example 1 in that LightGBM is replaced with a Bayesian model (BayesA model).

[0125] The prediction accuracy of Examples 1-4 and Comparative Example 1 was statistically analyzed, and the results are as follows: Figure 3 As shown in the figure, the accuracy of LightGBM as a classification model is significantly better than other models.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting a target trait suitable for low-depth sequencing, characterized in that, Includes the following steps: a. Obtain phenotypic data and whole-genome sequencing data of the samples; b. Genotype filling is performed on the whole genome sequencing data to obtain genotype data, and SNP loci associated with the target trait are screened from the genotype data; c. Using the SNP sites as features, train a classification model using the phenotypic data and genotypic data to obtain a trained classification model; d. Genotyping is performed on the whole-genome sequencing data of the sample to be tested, and then the trained classification model is used for classification to predict the target trait of the sample to be tested; The sequencing depth of the whole genome sequencing data was 0.5×-2×.

2. The method for predicting the target trait according to claim 1, characterized in that, In step b, a reference panel is constructed using GLIMPSE2 (2.0.0) software for the high-depth sequencing data, and then the whole genome sequencing data is filled with genotypes based on the constructed reference panel.

3. The method for predicting the target trait according to claim 1, characterized in that, During the genotype filling process, a hidden Markov model algorithm is used to infer the genotype of each sample at each locus.

4. The method for predicting the target trait according to claim 1, characterized in that, In step b, genome-wide association analysis is used to screen for SNP loci associated with the target trait.

5. The method for predicting the target trait according to claim 1, characterized in that, The classification models include LightGBM, random forest, deep neural network, support vector machine, or Bayesian model.

6. A genome-wide selection model suitable for low-depth sequencing, characterized in that, It includes an acquisition module, a preprocessing module, and a classification module; The acquisition module is used to acquire whole genome sequencing data of the sample to be tested; The preprocessing module is used to fill the genotypes in the whole genome sequencing data of the sample to be tested; The classification module is used to input the data processed by the preprocessing module into the pre-trained classification model, and perform classification processing through the classification model to predict the target trait of the sample to be tested; The classification model was trained using the following method: a. Obtain phenotypic data and whole-genome sequencing data of the samples; b. Genotype filling is performed on the whole genome sequencing data to obtain genotype data, and SNP loci associated with the target trait are screened from the genotype data; c. Using the SNP sites as features, train a classification model using the phenotypic data and genotypic data to obtain a trained classification model; The sequencing depth of the whole genome sequencing data was 0.5×-2×.

7. The genome-wide selection model according to claim 6, characterized in that, In step b, a reference panel is constructed using GLIMPSE2 (2.0.0) software for the high-depth sequencing data, and then the whole genome sequencing data is filled with genotypes based on the constructed reference panel.

8. The genome-wide selection model according to claim 6, characterized in that, During the genotype filling process, a hidden Markov model algorithm is used to infer the genotype of each sample at each locus.

9. The genome-wide selection model according to claim 6, characterized in that, In step b, genome-wide association analysis is used to screen for SNP loci associated with the target trait.

10. The genome-wide selection model according to claim 6, characterized in that, The classification models include LightGBM, random forest, deep neural network, support vector machine, or Bayesian model.

Images