A Machine Learning Method for Identifying Duck Breeds Based on Structural Variation
By combining structural variation with machine learning algorithms, the problem of low accuracy in duck breed identification has been solved, achieving efficient and accurate duck breed identification, especially providing a new technical means to distinguish breeds with similar genetic backgrounds.
Patent Information
- Application Number
- CN202611164404.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-03
- Publication Date
- 2026-11-17
AI Technical Summary
Existing duck breed identification methods suffer from low accuracy and difficulty in distinguishing breeds with similar genetic backgrounds.
By combining structural variation with machine learning algorithms, multiple structural variation detection software are used to jointly predict and find the intersection. Population genetic indicators are integrated to rank the information content, and a machine learning classification model is constructed. Rigorous hierarchical nested cross-validation is used for duck breed identification.
It improves the accuracy and stability of duck breed identification, and enables effective differentiation of breeds with similar genetic backgrounds, with an accuracy rate of up to 98.22%.
Smart Images

Figure CN123120048A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of poultry breed identification technology, specifically a machine learning method for identifying duck breeds based on structural variations. Background Technology
[0002] With the development of molecular breeding technology, molecular marker-based methods for identifying biological varieties have been widely used in fields such as genetic resource protection, variety authenticity identification, and germplasm resource evaluation. Currently, commonly used variety identification methods mainly include morphological identification, microsatellite marker (SSR) identification, and single nucleotide polymorphism (SNP) identification. However, morphological identification is easily affected by environmental factors, resulting in low stability and accuracy. SSR marker detection throughput is limited, and although SNP markers have advantages such as high detection efficiency and ease of standardization, the genetic information carried by a single locus is limited, making it difficult to achieve high-precision differentiation between varieties with similar genetic backgrounds.
[0003] Structural variants (SVs) are genomic variations that are typically longer than 50 bp, including types such as deletions, duplications, inversions, and translocations. Compared with SNPs, structural variants cover a larger genomic region, carry richer genetic information, and have a more significant impact on gene expression and phenotype formation. Therefore, they have great application potential in the field of variety identification.
[0004] Existing technologies construct classification models by screening characteristic SNP loci to identify poultry breeds. However, this method still suffers from the problem of limited information content per locus. The non-patent literature "Comprehensive analysis of structural variants in chickens using PacBio sequencing" reports the application of structural variations in chicken population genetic analysis and breed classification, indicating that structural variations have the feasibility of serving as molecular markers for breed identification.
[0005] However, there is currently a lack of technical solutions for efficient duck breed identification using structural variation combined with machine learning algorithms. Therefore, it is necessary to establish a duck breed identification method based on structural variation genetic markers and machine learning models to improve the accuracy and stability of duck breed identification and provide technical support for duck genetic resource protection, breed identification and molecular breeding. Summary of the Invention
[0006] The purpose of this invention is to provide a machine learning method for identifying duck breeds based on structural variation, in order to solve the problems of low accuracy and difficulty in distinguishing breeds with similar genetic backgrounds in the existing duck breed identification methods mentioned in the background art.
[0007] To achieve the above objectives, the present invention provides the following technical solution: a machine learning method for identifying duck breeds based on structural variations, the learning method comprising the following steps: S1. First, individual samples from multiple duck breeds need to be collected by cold collection. Whole-genome resequencing is performed on each sample to obtain raw sequencing data. The raw sequencing data is then subjected to quality control to remove low-quality reads and obtain clean reads. The clean reads are then aligned to the duck reference genome to obtain alignment results. At least two structural variation detection software programs are used to independently detect structural variations in the alignment results. The detection results from each software program are then merged and their intersection is taken to obtain a high-confidence structural variation dataset.
[0008] S2. Then, it is necessary to perform quality screening on the high-confidence structural variation dataset, remove low-quality structural variation sites and low-quality samples, obtain a high-quality structural variation dataset, calculate at least one population genetic index, sort the information content of each structural variation site in the high-quality structural variation dataset, and screen candidate structural variation marker sets of different sizes according to the sorting results.
[0009] S3. Finally, based on the candidate structural variant marker set, at least one machine learning classification model needs to be constructed, and hierarchical nested cross-validation is used to systematically evaluate each model. Based on the evaluation results, the optimal structural variant marker set and the optimal classification model are determined. The optimal classification model is then used to identify the breed of the duck sample to be tested, and the breed classification result is output.
[0010] Preferably, Fastp is used for quality control, BWA-MEM2 is used to align clean reads to the duck reference genome, and SAMtools is used to sort, remove duplicate sequences, and create an index for the generated BAM file.
[0011] Preferably, the at least two structural variation detection software programs include Smoove and Manta. SURVIVOR is used to merge the detection results of the two software programs and take the intersection. The merging parameters are set as follows: the maximum breakpoint spacing is 1000bp, at least two software programs are required to detect the structural variation simultaneously and the structural variation type and chain direction are consistent, and the minimum structural variation length is set to 50bp.
[0012] Preferably, BCFtools is used to screen the structural variation dataset after the intersection, retaining precise structural variations on autosomes that are biallelic and range in length from 50bp to 5Mb. PLINK is used to screen for structural variation sites with a minimum allele frequency of not less than 0.02, a genotype deletion rate of not more than 0.01, and individuals with a deletion rate of not more than 10%.
[0013] Preferably, the population genetic indicators include at least one of the following: population differentiation index (FST), mean difference between pairs of allele frequencies (Delta), variance spread distance (AED) between allele frequencies, and polymorphism information content (PIC). All structural variation sites are sorted in descending order according to the values of each indicator. Based on the sorting results, the top-ranked structural variations are selected to construct marker sets of different sizes. The marker quantity gradient includes 10, 20, 50, 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 1200, 1400, 1600, 1800, 2000, 2500, and 3000.
[0014] Preferably, the machine learning classification model includes at least one of Random Forest (RF), Linear Kernel Support Vector Machine (SVM_linear), Radial Basis Kernel Support Vector Machine (SVM_RBF), Naive Bayes (NB), Linear Discriminant Analysis (LDA), K-Nearest Neighbors (KNN), Decision Tree, XGBoost, LightGBM, Multilayer Perceptron (MLP), and Lasso Regularized Multiclass Regression.
[0015] Preferably, the specific process of the stratified nested cross-validation is as follows: all samples are randomly divided into K subsets according to variety as outer cross-validation. Each time, one subset is selected as the test set, and the remaining K-1 subsets are used as the training set. The population genetic index ranking results are independently recalculated within the training set. The classification model is trained based on the training fold data and prediction is performed on the independent test fold. This process is repeated K times so that each sample is tested once.
[0016] Preferably, the optimal classification model is a linear kernel support vector machine (SVM_linear), and the optimal structural variant marker set is the top 500 structural variant sites ranked by the polymorphic information content (PIC) index. The method further includes the step of constructing a genotype matrix for the sample to be tested: extracting genomic DNA from the duck sample to be tested, performing whole-genome resequencing, and after quality control, reference genome alignment, structural variant detection, and quality screening, extracting the genotype of each point in the optimal structural variant marker set, encoding and constructing a genotype matrix, inputting it into the optimal classification model for variety identification, and screening candidate variety identification biomarkers (BIMs) based on the multi-index ranking results, further analyzing the chromosome position, start coordinates, end coordinates, structural variant type, and structural variant length of the candidate structural variants to form a variety identification biomarker database.
[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention uses structural variations as genetic markers, which have higher information carrying capacity and genetic stability compared to traditional SNP markers, and can more effectively distinguish duck breeds with similar genetic backgrounds. This invention innovatively uses two structural variation detection software, Smoove and Manta, to jointly predict and take the intersection, to obtain a structural variation dataset with higher confidence, reducing the impact of false positive detection by a single software on downstream analysis.
[0018] This invention integrates four population genetic indicators—FST, Delta, AED, and PIC—to rank structural variations based on their information content. It also combines multiple machine learning classifiers for systematic benchmark testing and employs rigorous hierarchical nested cross-validation. The label ranking is independently recalculated within the training iteration to avoid information leakage from the test set, making the evaluation results more realistic and reliable.
[0019] This invention, after systematic evaluation, uses the PIC index to screen 500 structural variation markers and combines them with a linear kernel support vector machine classification model. The average recognition accuracy reached 95.02% (standard deviation 1.57%) with five-fold nested cross-validation. When the number of markers is further increased to 3000, the accuracy can reach up to 98.22% (AED index + linear kernel support vector machine). This invention establishes a duck breed identification process based on structural variation, providing a new technical means for duck genetic resource protection, breed authenticity identification, and molecular breeding. Attached Figure Description
[0020] Figure 1 This is a map showing the SV genome distribution of the present invention; Figure 2 This is a diagram showing the distribution of SV chromosomes in this invention. Figure 3 These are ten machine learning accuracy graphs from the present invention; Figure 4The confusion matrices of the SVM of the present invention under different numbers of sites are: (a) 100 SVs; (b) 200 SVs; (c) 400 SVs; (d) 1000 SVs. Detailed Implementation
[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below. 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.
[0022] Example 1: Materials and Methods Experimental instruments and equipment This invention mainly uses instruments and equipment such as PCR amplification instrument, real-time quantitative PCR instrument, NanoDrop spectrophotometer, Covaris ultrasonic disruptor, electrophoresis apparatus, agarose horizontal electrophoresis tank, TruSeq library, and Qubit fluorometer.
[0023] Experimental reagents and kits This invention mainly uses reagents and kits such as dsDNA fluorescence quantitative reagent kit, qPCR quantitative reagent kit, A-tailing reagent, adapter ligation reagent, and 1% agarose precast gel.
[0024] laboratory animals The sequencing data of Beijing ducks, Donglan ducks, and some Longsheng jade ducks, Xilin ducks and Rongshui ducks used in this invention are from public databases, while the remaining data were generated by blood sequencing in this experiment.
[0025] Experimental methods Domestic duck sample collection This invention utilizes whole-genome resequencing data from 348 domestic ducks, representing 22 native Chinese breeds. Of these, 318 individuals passed quality screening and were retained for subsequent analysis. Sequencing data for Beijing ducks, Donglan ducks, and some Longsheng Emerald ducks, Xilin ducks, and Rongshui ducks were obtained from public databases, while the remaining data were obtained through blood sequencing in this experiment.
[0026] DNA extraction and whole-genome resequencing Genomic DNA was extracted from blood samples, and its degradation and contamination were assessed using 1% agarose gel electrophoresis. DNA purity was measured using a NanoDrop spectrophotometer, with an OD260 / 280 ratio between 1.8 and 2.0. DNA concentration was quantified using a Qubit fluorometer, and only DNA exceeding 1.5 μg was used for subsequent library construction. Qualified DNA was sonicated to fragments of approximately 350 bp on average using a Covaris sonicator, and sequencing libraries were constructed using the TruSeq library construction kit according to standard procedures, including end repair, A-tailing, adapter ligation, purification, and PCR amplification. Library quality was assessed using Qubit, Agilent insert size evaluation, and qPCR quantification. Libraries with a concentration greater than 2 nM were sequenced on the Illumina platform, with an average sequencing depth of 8.62 × 10⁻⁶.
[0027] WGS Data Quality Control and Comparison Fastp (v0.23.4) was used to perform quality control on the raw sequencing data to obtain high-quality clean reads. Then, BWA-MEM2 (v2.2.1) was used to align them to the Beijing duck T2T reference genome (GCA_047663525.1). After alignment, SAMtools (1.16.1) was used to sort, deduplicate, and index the BAM files.
[0028] Example 2: Structural Variation Detection and Annotation This invention uses two software programs, Smoove (v0.2.8) and Manta (manta-1.6.0_py27), to independently detect structural variations in the alignment results. Smoove integrates structural variation (SV) detection and annotation software such as Lumpy, SVTyper, Duphold, and SVTools. Lumpy (v0.2.13) is used for SV detection, SVTyper (v0.7.1) is used for genotyping, and Duphold (v0.2.1) is used for annotation. Manta independently detects structural variations and performs genotyping based on the split-read and discrete read-pair signals in the alignment results. Subsequently, SURVIVOR (v1.0.6) was used to merge the SV detection results from Smoove and Manta. The maximum breakpoint spacing was set to 1000 bp, and only SVs detected by both software (i.e., both Smoove and Manta detected them) were retained. SV types and chain directions were also required to be consistent. No additional processing was considered for repetitive sequences, and the minimum SV length was set to 50 bp (SURVIVOR merge parameters: 1000 2 1 1 050). The strict intersection of the two software results was taken to obtain a high-confidence structural variation dataset. After annotation, BCFtools (v1.16) was used to filter the SV dataset after intersection, retaining SVs that met the following parameters: (1) INFO / IMPRECISE = 0; (2) Biallelic SVs on autosomes; (3) SV length 50bp < SV ≤ 5Mb; (4) INFO / MSHQ < 3, where MSHQ is the average SHQ score of all heterogeneous samples.
[0029] Exclude the following SVs: (1) DEL: DHFFC ≥ 0.7; (2) DUP: DHBFC ≤ 1.3.
[0030] Finally, PLINK (v1.90b6.21) was used to screen for SVs with MAF < 0.02, genotype deletion rate > 0.01, and deletion rate > 10%. The selected variants were used as the final SV dataset for subsequent analysis.
[0031] Table 1: Statistical Analysis of SV Type Characteristics
[0032] Example 3: Preprocessing of SV genotype data A structural variant (SV) genotype matrix was constructed using the structural variant genotype file (.raw format) exported from PLINK. Each sample corresponds to one row in the matrix, and each SV locus corresponds to one column. The genotype codes are 0, 1, and 2, representing the reference homozygous, heterozygous, and variant homozygous types, respectively.
[0033] The genotype information of the samples was extracted from the FID field in the file and converted into a breed vector. Missing values in the SV genotype matrix were filled with heterozygous genotypes (1) to ensure the integrity of subsequent machine learning model training. Finally, the SV genotype matrix (samples × SVs), genotype vector, and list of SV loci names were obtained and saved as an R object for subsequent analysis.
[0034] Example 4: SV information content ranking based on population genetic indicators To screen candidate structural variation markers that can effectively distinguish different varieties, this invention calculates four population genetic indicators: FST (population differentiation index), Delta (mean difference between pairs of allele frequencies among varieties), AED (expanded distance of variance of allele frequencies among varieties), and PIC (polymorphism information content). The information content of all SV loci is ranked and used as the basis for constructing the marker panel.
[0035] FST was calculated using the Weir & Cockerham (1984) unbiased estimation method, based on the ratio of between-group variance to within-group variance of allele frequencies for each variety. Delta was defined as the mean of the absolute values of the differences in allele frequencies between any two varieties. AED was calculated based on the variance expansion of allele frequencies for each variety relative to the overall mean. PIC was calculated based on the overall allele frequencies of the locus. All four indicators can be independently recalculated using only the training set data, avoiding the introduction of test set information during the label selection stage. All SVs were sorted in descending order according to the PIC indicator values to obtain a ranking list of the importance of structural variations.
[0036] Example 5: Constructing Tag Sets of Different Sizes Based on SV Importance Ranking Based on the ranking results obtained from FST / Delta / AED / PIC, the top-ranked SVs are selected in turn to construct marker panels of different sizes.
[0037] The set marker count gradient includes: [\ 10, 20, 50, 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 1200, 1400, 1600, 1800, 2000, 2500, 3000 ] For each label size (N) and each ranking metric, the top (N) SVs ranked by that metric are selected as the feature set to construct the corresponding SV label panel.
[0038] Example 6: Machine Learning Classification Model Construction and Performance Evaluation To evaluate the variety identification capabilities of SV marker sets of different sizes, this invention employs eleven supervised learning algorithms for classification modeling: Random Forest (RF), Support Vector Machine (Linear Kernel, SVM_linear), Support Vector Machine (Radial Basis Kernel, SVM_RBF), Naive Bayes Classifier (NB), Linear Discriminant Analysis (LDA), K-Nearest Neighbors (KNN), Decision Tree, XGBoost, LightGBM, Multilayer Perceptron (MLP), and Lasso Regularized Multiclass Regression.
[0039] (1) Random Forest (RF) A random forest classification model was constructed using 500 decision trees (ntree = 500) to evaluate the classification ability of the SV tag set.
[0040] (2) Support Vector Machine - Linear Kernel (SVM_linear) A support vector machine classifier is constructed using a linear kernel function to find the optimal classification hyperplane to distinguish between different varieties.
[0041] (3) Support Vector Machine - Radial Basis Kernel (SVM_RBF) A support vector machine classifier is constructed using a radial basis function kernel (cost = 1) to handle the case where varieties are non-linearly separable.
[0042] (4) Naive Bayes Classifier (NB) A classification model is established based on Bayes' theorem, assuming that each SV site is independent of the others.
[0043] (5) Linear Discriminant Analysis (LDA) Linear discriminant analysis (LDA) was employed to classify varieties by maximizing between-group variance and minimizing within-group variance. SV loci with zero variance in the training set were removed before LDA modeling to avoid singular covariance matrices.
[0044] (6) K-Nearest Neighbors (KNN) The K-nearest neighbor algorithm (K = 5) is used for classification, and the variety is determined based on the distance between the test sample and the training sample in the feature space.
[0045] (7) Decision Tree A single classification tree model was constructed using the CART decision tree algorithm (rpart, method = "class") as one of the baseline classifiers.
[0046] (8) XGBoost A classification model was constructed using the gradient boosting tree algorithm (nrounds = 100, max_depth = 6, eta = 0.3, objective = "multi:softmax").
[0047] (9) LightGBM A classification model is constructed using a gradient boosting algorithm based on histograms (num_leaves = 31, learning_rate = 0.1, nrounds = 100, objective = "multiclass").
[0048] (10) Multilayer Perceptron (MLP) A classification model is constructed using a single hidden layer neural network (size = 10, decay = 5e-4, maxit = 300, softmax output).
[0049] (11) Lasso Regularized Multiclass Regression L1 regularized (alpha = 1) multi-class logistic regression implemented using glmnet automatically determines the optimal regularization parameter lambda through 5-fold internal cross-validation (cv.glmnet).
[0050] Example 7: Rigorous hierarchical nested cross-validation All ranking metrics (FST / Delta / AED / PIC) and classification models are combined using rigorous hierarchical nested cross validation to evaluate classification performance, thus completely avoiding information leakage during the label selection phase.
[0051] Specifically, all samples are first stratified and randomly divided into 5 subsets by variety (outer layer cross-validation). Each time, one subset is selected as the test set, and the remaining 4 subsets are used as the training set. Within the training set, the FST / Delta / AED / PIC ranking results are recalculated independently to avoid using test set information for label selection. Then, each classification model is trained based on the training fold data and predictions are made on independent test folds. This process is repeated 5 times, with each sample tested once. The training and evaluation of all combinations of "ranking metric × number of labels × classifier" are performed in parallel in a multi-core environment using foreach / doParallel.
[0052] Classification accuracy is defined as:
[0053] in: (N_{correct}) represents the number of correctly classified samples; (N_{total}) represents the total number of test samples.
[0054] For each label size and each classification model, calculate the mean accuracy and standard deviation (SD) of the accuracy obtained from five cross-validations:
[0055] Finally, the performance of each classifier under different numbers of SVs was obtained.
[0056] Table 2: Recognition accuracy of each classifier under PIC index ranking (20-label panel size).
[0057]
[0058] Example 8: Determination of the optimal SV marker set and screening of candidate variety identification markers (Breed-Informative Markers, BIMs) Based on the classification accuracy trends corresponding to different ranking metrics and different numbers of SVs, plot the relationship curve of "number of labels - classification accuracy", and calculate the minimum number of labels required for each combination to reach the accuracy thresholds of 90%, 95%, and 99%.
[0059] Considering the balance between classification performance and the number of markers, this invention determines that the optimal combination for economical classification is the PIC index ranking combined with a linear kernel support vector machine (SVM_linear) model: with 500 markers, the average recognition accuracy of five-fold nested cross-validation reaches 95.02% (standard deviation 1.57%); if the number of markers is further increased to 3000, the accuracy can reach up to 98.22% (AED index + linear kernel SVM combination). Finally, the minimum number of SV core markers is determined to be 500, which is used for subsequent variety identification and biomarker screening analysis.
[0060] Based on the ranking results of the PIC indicators, candidate variety identification markers (BIMs) were selected from the SVs with the highest importance.
[0061] Further analyze the chromosomal location, start coordinates, end coordinates, and structural variation type (such as DEL, DUP, INS, INV, etc.) of candidate SVs, and calculate the SV length: [ SVLEN = End - Start ] The final database of candidate BIMs includes: SV Number (SV ID) Chromosome Starting position (Start) End position Structural Variation Type (SV Type) Structural variation length (SV Length) Information Content Score (PIC).
[0062] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A machine learning method for identifying duck breeds based on structural variation, characterized in that, Its learning method includes the following steps: S1. First, individual samples from multiple duck breeds need to be collected by cold collection. Whole-genome resequencing is performed on each sample to obtain raw sequencing data. The raw sequencing data is then subjected to quality control to remove low-quality reads and obtain clean reads. The clean reads are then aligned to the duck reference genome to obtain alignment results. At least two structural variation detection software programs are used to independently detect structural variations in the alignment results. The detection results from each software program are then merged and their intersection is taken to obtain a high-confidence structural variation dataset. S2. Then, it is necessary to perform quality screening on the high-confidence structural variation dataset, remove low-quality structural variation sites and low-quality samples, obtain a high-quality structural variation dataset, calculate at least one population genetic index, sort the information content of each structural variation site in the high-quality structural variation dataset, and screen candidate structural variation marker sets of different sizes according to the sorting results. S3. Finally, based on the candidate structural variant marker set, at least one machine learning classification model needs to be constructed, and hierarchical nested cross-validation is used to systematically evaluate each model. Based on the evaluation results, the optimal structural variant marker set and the optimal classification model are determined. The optimal classification model is then used to identify the breed of the duck sample to be tested, and the breed classification result is output.
2. The machine learning method for identifying duck breeds based on structural variation according to claim 1, characterized in that: The process involves using Fastp for quality control, using BWA-MEM2 to align clean reads to the duck reference genome, and using SAMtools to sort, remove duplicate sequences, and index the generated BAM files.
3. The machine learning method for identifying duck breeds based on structural variation according to claim 1, characterized in that: The at least two structural variation detection software programs include Smoove and Manta. SURVIVOR is used to merge the detection results of the two software programs and take the intersection. The merging parameters are set as follows: the maximum breakpoint spacing is 1000bp, at least two software programs are required to detect the same structural variation type and chain direction, and the minimum structural variation length is set to 50bp.
4. The machine learning method for identifying duck breeds based on structural variation according to claim 1, characterized in that: The BCFtools dataset of structural variants after intersection was used to filter the data, retaining precise structural variants on autosomes that are biallelic and range in length from 50bp to 5Mb. PLINK was used to filter structural variant sites with a minimum allele frequency of not less than 0.02, a genotype deletion rate of not more than 0.01, and individuals with a deletion rate of not more than 10%.
5. The machine learning method for identifying duck breeds based on structural variation according to claim 1, characterized in that: The population genetic indicators include at least one of the following: population differentiation index (FST), mean difference between pairs of allele frequencies (Delta), variance spread distance (AED) between allele frequencies, and polymorphism information content (PIC). All structural variation sites are sorted in descending order according to the values of each indicator. Based on the sorting results, the top-ranked structural variations are selected to construct marker sets of different sizes. The marker quantity gradient includes 10, 20, 50, 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 1200, 1400, 1600, 1800, 2000, 2500, and 3000.
6. The machine learning method for identifying duck breeds based on structural variation according to claim 1, characterized in that: The machine learning classification model includes at least one of Random Forest (RF), Linear Kernel Support Vector Machine (SVM_linear), Radial Basis Kernel Support Vector Machine (SVM_RBF), Naive Bayes (NB), Linear Discriminant Analysis (LDA), K-Nearest Neighbors (KNN), Decision Tree, XGBoost, LightGBM, Multilayer Perceptron (MLP), and Lasso Regularized Multiclass Regression.
7. The machine learning method for identifying duck breeds based on structural variation according to claim 1, characterized in that: The specific process of the hierarchical nested cross-validation is as follows: all samples are randomly divided into K subsets according to variety as outer cross-validation; each time, one subset is selected as the test set, and the remaining K-1 subsets are selected as the training set. Within the training set, the population genetic index ranking results are independently recalculated. The classification model is trained based on the training fold data and prediction is performed on the independent test fold; this is repeated K times so that each sample is tested once.
8. The machine learning method for identifying duck breeds based on structural variation according to claim 1, characterized in that: The optimal classification model is a linear kernel support vector machine (SVM_linear), and the optimal structural variant marker set is the top 500 structural variant sites ranked by the polymorphic information content (PIC) index. The method also includes the step of constructing a genotype matrix for the sample to be tested: extracting genomic DNA from the duck sample to be tested, performing whole-genome resequencing, and after quality control, reference genome alignment, structural variant detection, and quality screening, extracting the genotype of each point in the optimal structural variant marker set, encoding and constructing a genotype matrix, inputting it into the optimal classification model for variety identification, and screening candidate variety identification biomarkers (BIMs) based on the multi-index ranking results, further analyzing the chromosome position, start coordinates, end coordinates, structural variant type, and structural variant length of the candidate structural variants to form a variety identification biomarker database.
