Identification of Chicken Populations by Machine Learning Models Using the Minimum Number of SNPs

doi:10.21203/rs.3.rs-95706/v1

Download PDF

Research

Identification of Chicken Populations by Machine Learning Models Using the Minimum Number of SNPs

https://doi.org/10.21203/rs.3.rs-95706/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

A marker combination capable of classifying a specific chicken population could improve commercial value by increasing consumer confidence with respect to the origin of the population. This would also facilitate the protection of genetic resources, especially in developing countries.

Methods

In this study, a total of 20 lines 283 samples which were consist of Korean native chicken, commercial native chicken, and commercial broilers with layer population were used for finding the minimum number of marker combinations through the 600k high-density single nucleotide polymorphism (SNP) array. Application of the machine learning algorithms, a genome-wide association study (GWAS), linkage disequilibrium (LD) analysis, and principal component analysis (PCA) were used to distinguish a target (case) group from control chicken groups. In the verification of the selected markers, a total of 12 lines 182 samples were used to confirm the change in the accuracy of the target chicken breed identification.

Results

A total of 47,303 SNPs was used for classifying chicken populations; 96 LD-pruned SNPs (50 SNPs per LD block) served as the best marker combination for target chicken classification. Moreover, 36, 44, and 8 SNPs were selected as the minimum numbers of markers by Adaboost (AB), Random Forest (RF), and Decision Tree (DT) machine learning classification models, which had accuracy rates of 99.6%, 98.0% and 97.9%, respectively. The selected marker combinations increased the genetic distance between the case and control groups, and reduced the number of genetic components, confirming that an efficient classification of the groups was possible using small number of marker sets. In a verification study including additional chicken breeds and samples, the accuracy did not significantly change, and the target chicken group could be clearly distinguished from the other populations.

Conclusions

The GWAS and PCA analysis, machine learning algorithm used in this study is able to be applied efficiently to explore the minimum combination of markers that can distinguish varieties among a large number of SNP markers.

Animal Science

Biotechnology and Bioengineering

single nucleotide polymorphism (SNP)

principal component analysis (PCA)

genome-wide association study (GWAS)

linkage disequilibrium (LD)

machine learning

Chicken is a rich source of protein in the human diet. The consumption of chicken has increased globally due to increased consumer interest in health and is also rising annually in Korea (Yeung & Morris 2001). It has been reported that Koreans consume about 2,347 tons of chicken every year, which equates to more than nine chickens per person (MAFRA, 2019). There has been a gradual shift in emphasis from price to quality, including taste and functionality, for meat products. Many chicken breeds with improved quality have been produced, but methods are required to certify them.

Generally, the methods used to identify chicken breeds are based on morphological features, but meat products already on the market cannot be identified by their morphological characteristics. A precise identification method allowing for verification at the genetic level is required. Animal genetic information can be used for maintenance and improvement of livestock varieties, based on phenotypes and heredity characteristics. The Korean government is currently developing genetic markers that can distinguish between cattle and pig breeds (Shim et al. 2010; Oh et al. 2014; Kim et al. 2015). These markers are microsatellite (MS) markers, with a high polymorphism of a single allele, but identifying genotypes requires many specialized personnel to perform polymerase chain reaction (PCR) and fragment analyses. Alternatives are needed to overcome these challenges (Dalvit et al. 2007).

Single nucleotide polymorphism (SNP) markers could be an important marker-based verification system, with the potential to replace MS markers. With the release of the draft genome sequence of the chicken in 2004, genome-wide SNPs have become available for various research applications. However, many restrictions have been placed on their use, due to the expense and the fact that the technology required to customize the desired SNPs is highly specialized (Burt 2005; Hillier et al. 2014). Recently, it is easy to create SNP kits to identify and validate chicken lines/strains developed for the quality of their meat, where various platforms can be used to combine SNPs in a similar manner to the kits used to diagnose diseases (Karniol et al. 2009; Futema et al. 2018).

Selecting SNP markers that can distinguish among livestock breeds and varieties is not easy. Unlike MS markers, SNPs are the same in all varieties/breeds, but normally possess one of three genotypes (AA, AB, or BB). The process of reducing the number of SNP markers for a particular population can affect the results. A fast and accurate SNP marker verification system running on an automated platform is needed as an alternative to existing verification systems based on MS markers (Vignal et al. 2002).

Machine learning is a type of artificial intelligence in which algorithms are developed that allow computers to make predictions through learning via training data (Mitchell 1999; Guinand et al. 2002). Predictions based on complex data structures (e.g., big data) that cannot be made by humans are the goal of machine learning. Data mining, string analysis, and pattern analysis can be performed using various classification models. Machine learning algorithms are rarely used in genetic diversity studies. However, a recent study compared classification performance, based on high-density SNPs, between support vector machine (SVM) and random forest (RF) models (Bertolini et al. 2015; Bertolini et al. 2017; Pasupa et al. 2020; Schiavo et al. 2020). Studies of machine learning algorithms for classification have not applied them to raw data, instead using filtered data, such as F statistic, delta statistic, or principal component analysis (PCA) eigenvalues, for classification. This helps machine learning algorithms to derive more efficient and accurate classification results, compared to the process of finding a combination of classification markers within an unspecified number of alleles (Bertolini et al. 2017; Judge et al. 2017).

The Korean chicken industry developed rapidly after the Korean War and became industrialized. However, most of the native chicken genetic resources were lost during the war, and the remaining populations were maintained only in small breeding families in backyards. They had low productivity and made little contribution to the overall development of the industry. The entire industry consisted of commercial broilers and layers. Thus, a project to restore native chicken breeds was launched in the 1990s by the Korean government and a few private companies that maintained small-scale native chicken populations (Jin et al. 2017). A project to develop a new chicken breeding stock was also initiated by the government, and a precise technique for the identification of new chicken breeding stocks at the genetic level was considered the best way to prevent the newly developed breed from being unfairly distributed in the market.

The purpose of this study was to identify the minimum number of SNP markers needed to identify and verify a target population from among other populations, using information obtained from a 600K SNP genotyping array for chicken.

Experimental animals and DNA extraction

Two sets of samples were used in this study. The initial set included a total of 283 samples (from 20 chicken populations) that were used for a high-density SNP array analysis. This analysis was performed to identify a combination of SNP markers capable of distinguishing the new chicken breeding stock. Samples in this set were divided into four groups: purebred Korean native chicken (KNC), commercial native chicken, commercial broiler, and commercial layer (Table 1). The purebred KNC population consisted of pure lines of KNC and adapted chicken lines, which had been preserved by the National Institute for Animal Science (NIAS, RDA, South Korea) (Seo et al. 2018). The second group included three commercial native chicken lines, including a founder group (HF, HH, and HY) that yielded a target group for breed identification. These three lines were maintained by a private company and used for commercial chicken (CC) production. The third and fourth groups, commercial broiler and layer lines, respectively, were used as comparison groups (Table 1). The second sample set consisted of 13 populations and 182 samples. Additional samples were included from the abovementioned populations, and a new commercial native chicken breed was used for validation of SNP combinations in the initial sample set. The detailed sample information for both sets is given in Table 1.

All samples used in this experiment were collected according to guidelines issued by the Institutional Animal Care and Use Committee of Chungnam National University, who approved this study (approval no. CNU-00486). Genomic DNA (gDNA) was extracted from whole blood samples taken from the wing vein of birds, using an EDTA-coated tube to prevent coagulation. Muscle tissue samples were obtained from chicken meat purchased from a market. The gDNA extraction was performed according to the manufacturer’s protocol using a PrimePrep™ genomic DNA isolation kit for blood and tissue (GeNetBio, Daejeon, South Korea). The quality and concentration of the extracted gDNA was verified with electrophoresis using 1% agarose gel, and spectroscopic analysis using a NanoDrop spectrophotometer (Thermo Fisher Scientific, Waltham, MA, USA).

High-density SNP genotyping and quality control (QC)

High-density SNP genotyping of the entire genome was performed using an Axiom 600K SNP genotyping array for chicken (Affymetrix, Santa Clara, CA, USA). A total of 580,954 genotypes were analyzed and the data were transformed into a binary file format using PLINK software (version 1.9; http://pngu.mgh.harvard.edu/~purcell/plink/download.shtml). The SNP results were subjected to a QC procedure, with the two main criteria of genotype error (< 90%) and minor allele frequency (<0.01) used for the selection of SNP markers in genetic diversity analyses. After the QC procedure, a population structure analysis was performed, based on a multidimensional scale (MDS) plot and admixture analysis, to identify similarities and differences between the target population and the other chicken populations. The MDS plot obtained with PLINK was used to analyze information on pair-wise genetic distances via a four-dimensional scale (Chang et al. 2015). The genetic components of each population were analyzed using ADMIXTURE software (version 1.3); the distributions of the genetic components of the populations were compared according to the numbers of random common ancestors based on the optimum K value (Alexander & Lange 2011). The results of the two analyses were represented graphically by a scatterplot and bar graph using R software (Team 2013).

Selection of 96 candidate SNP markers for identification of the target population

A detailed summary of the process used for the selection of candidate SNP markers distinguishing the new chicken breeding stock (with HH, HF, and HY as the parental lines) is provided in Figure 1. We used two main strategies to select a marker combination that distinguished the new chicken breeding stock. In the first step, SNPs were selected using the case/control association analysis in PLINK 1.9. In this analysis, the new chicken breeding stock was the case group and the other populations were the control group. P-values were derived for each SNP (Chang et al. 2015). SNPs were mainly identified in the macro-chromosome, indicating marker selection bias. In the second step, population linkage disequilibrium (LD) was analyzed using the significant SNPs obtained in step 2, to identify SNP markers that were evenly distributed throughout the entire genome. Three sets of 96 significant SNP marker combinations were thus obtained. The accuracy of the classification was compared among three scenarios. In the first scenario, SNPs with significantly lower p-values in a genome-wide association study (GWAS) association test were selected. In the second scenario, 1 SNP per LD block was selected. In the third scenario, 50 SNPs per LD block were selected. An MDS plot of each scenario was constructed to show the degree of separation of the target group from the other populations. In addition, custom SNP assays were designed for verification using additional chicken samples: a total of 182 samples from 13 populations were collected and genotyped by a Fluidigm array (Fluidigm, San Francisco, CA, USA) (Table 1).

Machine learning approach for breed identification

The 96 selected SNPs, which were identified with the 600K SNP genotyping array for chicken, were used as the training dataset. The data obtained via a verification study with the Fluidigm assay were used as the test dataset, with the target population identified using classification algorithms of machine learning techniques. We applied eight models to classify varieties/breeds: Random Forest (RF), Adaboost (AB), quadratic discrimination analysis (QDA), Naïve Bayes, nearest neighbor classification, linear discrimination analysis (LDA), and decision tree (DT) classification. Each machine learning model had its own criterion for determining whether the target population was consistently classified (Breiman 2001; Kégl 2013; Singh et al. 2016; Tharwat 2016). The sensitivity refers to the proportion of positive values that were accurately determined, i.e., the true-positive rate (TPR):

The specificity refers to the proportion of negative values that were accurately determined, i.e., the true-negative rate (TNR):

Where, TP means true-positive, TN means true-negative, FN means false-negative, and FP means false-positive (Altman & Bland 1994).

The HH, HF, and HY populations were shown by crossbreeding tests to be the best combinations for producing new chicken breeding stocks (data not shown). We sought the minimum number of marker combinations required to classify these three founder populations, and the great grandparent (GGP), grandparent (GP) and CCs produced through their mating. The overall procedure for this is shown in Figure 1.

Genetic diversity analyses to identify SNP marker combinations

To identify the target chicken population among the 20 populations included in this study, genetic clustering was performed. The genetic components of each population were confirmed through the MDS plot and admixture analysis. The MDS plot showed that PC1 and PC2 explained 44.414% of the total variance. The HH and HF founder groups were clustered together, directly under the clusters of commercial broiler groups (Cobb, Ab, and ROSS; Figure 2a). HA was also close to the ROSS and Cobb populations. In contrast, HY was in a separate cluster that included the HS, HW, NC, and ND populations. The commercial layer populations, HL and LO, were located in the -0.05 region of PC1. The most central clusters were identified as purebred KNCs from the NR and NY populations, which were clustered with NH and NS. These are known as the Cornish breed and were located in adjacent regions of the plot. In addition, HZ was confirmed to be a similar breed to NH. The HG and HV breeds formed the most independent cluster among all populations in this study. In the PCA plot shown in Figure 2B, PC1 and PC3 explained 38.864% of the total variance. HH, HF, and HY were distributed in different areas of the plot from the commercial broiler population and formed distinct clusters from the other breeds (Figure 2B).

The results of a genetic distance analysis and the fixation index (Fst), calculated based on the 96 SNPs selected from the 600K SNP genotyping array, are shown in Figure 3A. The results were consistent with those of the MDS plot (Figure 2). The HH and HF founder populations were genetically close to the commercial broilers of the Cobb (0.086 and 0.096), Ab (0.097 and 0.107), and ROSS (0.086 and 0.095) breeds (Figure 3A). Both of these founder populations were related to meat-type chicken breeds, and HZ (0.065) was also close to these populations. The Fst results confirmed that the genotype frequency was the same between the HH and HF founder populations and meat-type chicken breeds (0.138~0.175). The HY population was closest to the HS (0.093) and HW (0.077) clusters, and these breeds were closer to the commercial layer populations of LO (0.126) and HL (0.129) than the other chicken populations. The Fst confirmed that HY shared genotypes with HS and HW (Figure 3B).

The admixture results for the 20 chicken populations were used to compare the genetic components among the groups. The lowest cross-validation (CV) error was found at K=13 (Figure 4A; Supplementary Figure 1A). HH, HF, and HY, which were used as the founder populations for the new chicken breeding stock, had independent genetic components, although the HH and HF populations also shared some genetic components. It was also confirmed that the Ab, Cobb, ROSS, and HZ chicken populations had similar genetic components. Similar to the MDS results, the HL and LO commercial layer populations had the same genetic components; the Rhode Island Red breeds, NC and ND, also shared genetic components (Figure 4A). The founder populations of the new chicken breeding stock (HH, HF, and HY) had different genetic components. The KNCs (NR, NY) shared some of genetic components with other chicken breeds, such as the Cornish breeds (NS, NH), shown by their central location in the MDS plot (Figure 2).

GWAS and SNP selection

A GWAS was performed to identify the founder populations of the new chicken breeding stock and the other chicken populations: 47,303 SNPs were used to distinguish between the populations (Supplementary Figure 2). These markers represented about 10% of the 600K SNP genotyping array data and were obtained by applying Bonferroni correction to the GWAS analysis results, in which the p-value for significance was 0.05. LD pruning of the case and control groups was then performed to select an even number of markers from each chromosome to distinguish the target group. The discriminatory power of the marker combinations was compared with the PCA results. It was confirmed that scenario 3 (SNPset3) could efficiently distinguish between the different chicken groups (Figure 5C). In scenario 1 (SNPset1), PCA involved selection only of the most significant SNPs in the GWAS. In scenario 2 (SNPset2), in which 1 SNP per LD block was selected, the case and control groups were not clearly distinguished. It was confirmed that 95.8% of the SNPs in SNPset1, 72.9% of the SNPs in SNPset2, and 38.5% of the SNPs in SNPset3 were distributed in the largest chicken chromosome, GGA1 (Supplementary Figure 3).

Breed identification by machine-learning algorithms

We used the eight machine learning classification models to classify the founder group based on the 96 selected markers: more than 98.5% of the case and control samples were distinguished in all models (Figure 6). All models except the naïve Bayes one had 100% identification power, in terms of the sensitivity to confirm TPs and specificity to confirm FPs. The Adaboost (AB), Decision Tree (DT), and Random Forest (RF) algorithms also sought a solution involving the fewest markers: breed classification was possible with 36 markers for AB (Figure 7A), 44 markers for RF (Figure 7B), and 8 markers for DT (Figure 7C).

Validation study using additional samples

Additional samples were collected to verify the ability of the selected marker combinations to distinguish the founder population and their offspring. Samples of founder groups, commercial native chickens (WM2, Yelim K, HI), commercial broilers (Ab, Cobb, Ross), and commercial layers (LO) were collected from the Korean chicken market. To confirm the discrimination ability of CCs produced by crossing with the founder population, CCs from the HH, HF, and HY populations were used for verification (182 samples: Table 1).

The 96 selected markers from SNPset3 were genotyped by a customized Fluidigm assay. The case/control association results were almost the same before versus after adding the verification samples. When we selected the minimum number of SNP markers using the ‘feature selection’ function for the AB, DT, and RF algorithms, discriminatory power exceeding 99% for all three models using only the 283 samples in sample set1. In the verification study, the machine learning algorithm was trained using sample set1, and the case and control chicken populations were predicted using the sample set2 as the validation sample (Table 1). The target population was classified with 99.6%, 97.9% and 98.0% accuracy by the AB, DT, and RF algorithms, respectively (Table 2).

The ability to identify chicken breed or brands on the market at the genetic level could increase consumer trust. Previously used mtDNA sequence variation and MS markers remain useful to verify breeds. However, establishing an automated verification system for these methods take a long time, and an experienced operator with analytical skills is also required (Oh et al. 2014; Guo et al. 2017). SNP markers provide limited variant information compared to MS markers; however, a combination of several SNPs can provide sufficient information for classification. In addition, the cost of genotyping is continually falling, and customizable SNP genotyping platforms can be used as next-generation verification tools that can respond accurately and quickly to market demands.

However, identifying the minimum number of markers from a high-density SNP array for identification of a target population is not simple. In previous studies, independent SNPs determined by canonical discriminant analysis (CDA), the delta statistic, the F statistic, and PCA were used for genetic classification, and breed identification using low-density SNP arrays has also been demonstrated (Dimauro et al. 2013; Bertolini et al. 2017; Judge et al. 2017; Schiavo et al. 2020). In these studies, using an 600K SNP genotyping array for chicken, three combinations of 96 SNP markers were selected based on the results of a GWAS and LD analysis, where the new chicken breeding stock (with HH, HF, and HY as the founder populations) was the case and the remaining chicken groups were the controls. The feature selection function was applied SNPset3 to determine the minimum number of markers required for discrimination of the target group. The machine learning algorithm showed high discriminatory power (Figure 1).

Identification of target chicken population based on genetic components

New chicken breeding stocks produced by three-way crossing require a combination of shared markers that can be used to clearly distinguish them from the other chicken populations. Twenty chicken populations were used in this study (Figure 4). Of these, twelve chicken populations (HH and HF, HG and HV, HS and HW, NC and ND, NS and NH, NR and NY) had a shared origin; therefore, a total of 14 chicken populations was predicted as the independent chicken breeds. The HS, HW, NC, and ND lines all originated from Rhode Island Red (Seo et al. 2018), and the CC lines also shared part of genetic components with them. Twelve genetic components could be used to determine the origins of the chicken populations; it was difficult to discriminate them using fewer marker genotypes.

The populations to be classified had HH, HF, and HY as their parental lines. It was difficult to distinguish HH, HF, and HY from the other chicken populations using a limited number of SNP markers. In terms of genetic distance, HH and HF were very close (0.09), but HY was relatively distant from those two breeds on the MDS plot (genetic distances of 0.25 and 0.27, respectively). The HY population was more closely related to the other chicken populations than HH and HF. Therefore, it was difficult to identify a marker shared by all three founder populations. The same approach was used to classify breeds by population-specific alleles, similar to the existing mtDNA and MS marker classification approaches. However, different results were obtained using the different marker combinations when the verification samples were added (data not shown) because the SNPs extracted from the array were not conserved in each population. On the other hand, mutations in mtDNA or MS markers do not affect the function of genes, and are selected based on the mutation occurring from the maternal origin of the population (mtDNA marker), or the specific allele (MS marker) of the population is used as an identification point for classification. Thus, it was difficult to discriminate populations with a small number of samples using the population-specific SNPs.

GWAS and LD analysis for identification of the target population

Classification analysis was performed to overcome the limitations of the population-specific markers mentioned above. The HH, HF, and HY populations were set as the case group, and the remaining 17 populations were set as the control group. The 96 markers selected by the GWAS were strongly related to the case group. The case and control groups tended to form distinct clusters but were not clearly distinguished using only the GWAS significant SNPs. Therefore, LD pruning was performed and confirmed that SNPset3, which selected 50 SNPs per LD block, could clearly distinguish the two chicken groups due to remove the sharing of LD blocks between markers, or the relationship between adjacent LD blocks.

Regarding SNPset1, individuals with high genetic similarity had a high degree of clustering. Several samples overlapped in the MDS plot. When using SNPset2 and SNPset3, which selected SNPs based on the LD block, the clusters were separated by their relationships. If the SNP markers were selected based on their p-values in the GWAS, those having a strong correlation with the case group were affected by the LD relationships between marker distances. It was confirmed that 95.8% of SNPset1 (92 of 96 SNPs) shared 39 LD blocks on GGA1 (Supplementary Figure 3). Additionally, 70 SNPs in SNPset2, and 37 in SNPset3, were located on GGA1. Using the AdaBoost model, which had excellent discriminatory power, only six SNPs were selected from GGA1. Thus, many SNPs were strongly related to the case group in GGA1 but provided redundant information and probably interfered with the classification of the two groups. Selection of SNPs in the case group based on GWAS analysis could increase the genetic distance between the case and control groups. It was difficult to distinguish the two groups based on the Fst, but it was confirmed that the genetic distance between the case and control groups was significantly increased (Figure 3). The optimum K-value in the admixture analysis decreased from 12 to 2 at the minimum number of marker combinations by the Adaboost model (Figure 4).

In previous studies, methods for selecting the minimum number of high-density SNP markers for breed identification using the delta statistic and Fst were reported. More than 300 and 591 breed-specific SNPs were selected by Judge et al. (2017) and Kumar et al. (2019), respectively. These relatively large numbers of SNPs were used to form a panel to discriminate among target breeds. Another study sought to identify the minimum number of markers needed for breed identification using the delta statistic, PCA, and an RF algorithm (Bertolini et al. 2017). Combinations of markers (48- and 96-SNP panels) capable of distinguishing among various cattle breeds were presented; efficient identification was possible with fewer markers than in previous studies.

Our GWAS and LD analysis were not performed to identify markers capable of distinguishing among all of the populations included in the study, but rather to distinguish only the target population from the others. It is therefore difficult to directly compare the results with those of previous studies. Comparing the Fst, the genetic distance, and the genetic component of the research population before and after marker selection, it was confirmed that the changes of genetic distance and genetic composition calculated by the selected marker were significantly changed for target population except for Fst. The genetic distances were calculated based on allele frequencies, and the results were similar to those obtained using the delta score. The explanatory power of the principal component in the PCA analysis increased when using case-associated markers. The validation study results remained consistent after adding samples from other populations that were not used for marker selection. The 96 selected markers (SNPset3) well explained the genetic components of the target chicken group.

Machine learning algorithms for classification of the case and control chicken populations

Machine learning is a supervised learning approach for classifying new observations that can be used to classify bi-class or multi-class data. Machine learning can be used for voice and handwriting analysis, and document classification. In recent years, machine learning and deep learning algorithms have been used to determine phenotypic associations (e.g., in the genome, transcriptome, and methylome) in “omics” research, and to establish classification models (Pérez-Enciso & Zingaretti 2019; Alves et al. 2020).

In this study, eight machine learning classification models were used to efficiently identify target chicken populations. PCA was conducted with a machine learning algorithm to confirm whether case and control groups could be distinguished, based on the 96 markers in the SNPset3. All classification algorithms showed 100% breed classification accuracy except the Naïve Bayes model (98.5%; Figure 6). The AB, RF, and DT algorithms select a subset of variables through a feature selection process. In general, machine learning models utilize this method to: 1) simplify the model for easier interpretation, 2) shorten the training time, 3) avoid the dimensional curse problem, and 4) reduce overfitting (i.e., reduce variance) (Bermingham et al. 2015). In this manner, duplicate or less relevant variables are removed, so the minimal number of SNP markers required for efficiently classifying chicken populations can be identified.

With use of feature selection, 36, 44 and 8 SNP markers were selected by the AB, RF, and DT models, which had classification accuracies of 99.6%, 97.9% and 98.0%, respectively (Table 2). Thus, the target group could be classified using a small number of markers.

In the validation study including additional samples, both founder group and non-founder group chickens could be classified. The added samples included PL and CC samples from the founder population, and various samples obtained from the Korean chicken market (including new breeds not included in the 600K SNP genotyping array; e.g., WM_2, Yelim K, and HI). The discriminatory power was excellent, even when samples from a group of chickens with a completely different genetic composition to that used in the initial marker selection process were added (Figure 7). Our SNP markers were not population-specific. Therefore, by adding samples, the allele frequency and breed classification results could change. However, no significant changes were seen, and the cases and control groups were classified with high accuracy (Table 3).

The AB and RF models showed similar clustering results with and without the additional verification samples, while the DT model produced relatively diffuse clusters. Overall, the accuracy was similar among the classification models. The SNP marker combinations selected by the three models can be used for classifying the target chicken population. However, the DT model showed changes in clustering with additional samples, so requires further verification.

A genetic marker capable of distinguishing among breeds at the genetic level is required to protect intellectual property rights and ensure consumer confidence, but the development of conventional mtDNA and MS markers requires large amounts of time and money, as well as expertise. In this study, the minimum number of SNPs that could identify target chicken populations was determined. The marker combinations thus selected, where these markers were relatively distant in terms of their LD relationships, were effective for distinguishing the case and control chicken groups. Machine learning models suggested the most effective marker combinations by minimizing redundant marker information. Based on the GWAS and LD analysis, machine learning models could be efficient for selecting the minimum number of SNP markers needed for breed classification. The minimum number of marker combination developed in this study has commercial potential for chicken population identification, especially of the new KNC population.

AB: Adaboost

Ab: Avor Acre

CC: commercial chicken

CDA: canonical discriminant analysis

Cobb: Cobb broiler

CV: cross-validation

DT: Decision Tree

FN: false-negative

FP: false-positive

GGP: great grandparent

GP: grandparent

GWAS: genome-wide association study

HA: Hanhyup A

HF: Hanhyup F

HG: Hanhyup G

HH: Hanhyup H

HI: Hyun-in farm commercial native chicken

HL: Hy-line layer

HS: Hanhyup S

HV: Hanhyup V

HW: Hanhyup W

HY: Hanhyup Y

HZ: Hanhyup Z

KNC: Korean native chicken

LD: linkage disequilibrium

LDA: linear discriminant analysis

LO: Lohman brown layer

MDS: multi-dimensional scaling

MS: microsatellite

mtDNA: mitochondrial DNA

NC: NIAS Rhode Island Red C

ND: NIAS Rhode Island Red D

NH: NIAS Cornish H

NIAS: National Institute for Aniaml Science

NR: NIAS KNC Red-brown

NS: NIAS Cornish S

NY: NIAS KNC Yellow-brown

PCA: principal component analysis

PCR: polymerase chain reaction

PL: parental line

QDA: quadratic discriminant analysis

RF: Random Forest

ROSS: Ross broiler

SNP: single nucleotide polymorphism

SVM: support vector machine

TN: true-negative

TNR: true-negative rate

TP: true-positive

TPR: true-positive rate

WM2: Woorimatdaq ver2 commercial native chicken

Yelim K: Yelim farm commercial native chicken

Ethics approval and consent to participate

Consent for publication

All of the authors have approved the final version of the manuscript and agreed with this submission to the Journal of Animal Science and Biotechnology.

Availability of data and material

Datasets obtained and analyzed in this research are included within this article (and the supplementary data files).

Competing interests

The authors declare that they have no competing interests.

Funding

Author’s information

Affiliations

Division of Animal and Dairy Science, Chungnam National Univeristy, Daejeon 34134, Republic of Korea, Bio-AI Convergence Research Center, Chungnam National Univeristy, Daejeon 34134, Republic of Korea, SELS center, Division of Biotechnology, Advanced Institute of Environment and Bioscience, Chonbuk National University, Iksan 54596, Republic of Korea, Insilicogen inc. Yongin 16954, Republic of Korea

Contributions

DS, HYK, JHL designed the research. NRC performed experiments. DS, SHC, HYK analyzed, and interpreted the data. DS, PM, YKK, YJK, HYK, JHL and HL interpreted the results, edited, reviewed, and approved the final version of the manuscript. The authors read and approved the final manuscript.

Corresponding author

Correspondence to Hyeong Yong Kim and Jun Heon Lee

Acknowledgements

This work was supported by the Korea Institute Planning and Evaluation for Technology in Food, Agriculture, Forestry and Fisheries through the Golden Seed Project, Ministry of Agriculture, Food and Rural Affairs (2013010-05-4-SB250) and the “Cooperative Research Program for Agriculture Science & Technology Development (PJ0128202020)” of the Rural Development Administration, Republic of Korea, and partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT)(No.2020-0-01441, Artificial Intelligence Convergence Research Center(Chungnam National University)) and Korea Evaluation Institute of Industrial Technology(KEIT) grant funded by the Korea government(MOTIE).

Alexander DH, Lange K. Enhancements to the ADMIXTURE algorithm for individual ancestry estimation. BMC Bioinformatics 2011;12(1):246.
Altman DG, Bland JM. Diagnostic tests. 1: Sensitivity and specificity. BMJ Brit Med J. 1994;308(6943):1552.
Alves AAC, da Costa RM, Bresolin T, Fernandes Júnior GA, Espigolan R, Ribeiro AMF, et al. Genome-wide prediction for complex traits under the presence of dominance effects in simulated populations using GBLUP and machine learning methods. J Anim Sci. 2020.
Bermingham ML, Pong-Wong R, Spiliopoulou A, Hayward C, Rudan I, Campbell H, et al. Application of high-dimensional feature selection: evaluation for genomic prediction in man. Sci Rep-UK. 2015;5:10312.
Bertolini F, Galimberti G, Calò D, Schiavo G, Matassino D, Fontanesi L. Combined use of principal component analysis and random forests identify population‐informative single nucleotide polymorphisms: application in cattle breeds. J Anim Breed Genet. 2015;132(5):346-56.
Bertolini F, Galimberti G, Schiavo G, Mastrangelo S, Di Gerlando R, Strillacci M, et al. Preselection statistics and Random Forest classification identify population informative single nucleotide polymorphisms in cosmopolitan and autochthonous cattle breeds. Animal. 2017;12:12-9.
Breiman L. Random Forests. Mach Learn. 2001;45:5-32.
Burt DW. Chicken genome: current status and future opportunities. Genome Res. 2005;15(12):1692-8.
Chang CC, Chow CC, Tellier LC, Vattikuti S, Purcell SM, Lee JJ. Second-generation PLINK: rising to the challenge of larger and richer datasets. Gigascience 2015;4(1):s13742-015.
Dalvit C, De Marchi M, Cassandro M. Genetic traceability of livestock products: A review. Meat Sci. 2007;77(4):437-49.
Dimauro C, Cellesi M, Steri R, Gaspa G, Sorbolini S, Stella A, Macciotta NPP. Use of the canonical discriminant analysis to select SNP markers for bovine breed assignment and traceability purposes. Anim Genet. 2013;44(4):377-82.
Futema M, Bourbon M, Williams M, Humphries SE, Clinical utility of the polygenic LDL-C SNP score in familial hypercholesterolemia. Atherosclerosis. 2018;277:457-63.
Guinand B, Topchy A, Page K, Burnham-Curtis M, Punch W, Scribner K. Comparisons of likelihood and machine learning methods of individual classification. J Hered. 2002;93(4):260-9.
Guo H, Li C, Wang X, Li Z, Sun G, Li G, Liu X, Kang X, Han R. Genetic diversity of mtDNA D-loop sequences in four native Chinese chicken breeds. Brit Poult Sci. 2017;58(5):490-7.
Hillier LW, Miller W, Birney E, Warren W, Hardison RC, Ponting CP, et al. Sequence and comparative analysis of the chicken genome provide unique perspectives on vertebrate evolution. Nature. 2014;423(10):695-777.
Jin S, Jayasena D, Jo C, Lee J. The breeding history and commercial development of the Korean native chicken. World Poult Sci J. 2017;73(1):163-74.
Judge M, Kelleher M, Kearney J, Sleator R, Berry D. Ultra-low-density genotype panels for breed assignment of Angus and Hereford cattle. Animal. 2017;11(6):938-47.
Karniol B, Shirak A, Baruch E, Singrün C, Tal A, Cahana A, Kam M, et al. Development of a 25‐plex SNP assay for traceability in cattle. Anim Genet. 2009;40(3):353-6.
Kégl B. The return of AdaBoost. MH: multi-class Hamming trees. arXiv. 2013;1312.6086.
Kim K, Seo M, Kang H, Cho S, Kim H, Seo KS. Application of logitboost classifier for traceability using snp chip data. PLoS One. 2015;10(10):e0139685.
Kumar H, Panigrahi M, Chhotaray S, Parida S, Chauhan A, Bhushan B, Gaur G, Mishra B, Singh R. Comparative analysis of five different methods to design a breed-specific SNP panel for cattle. Anim Biotechnol. 2019:1-7.
Mitchell TM. Machine learning and data mining. Commun ACM. 1999;42(11):30-6.
Oh JD, Song KD, Seo JH, Kim DK, Kim SH, Seo KS, Lim HT, Lee JB, Park HC, Ryu YC. Genetic traceability of black pig meats using microsatellite markers. Asian Austral J Anim. 2014;27(7):926.
Pasupa K, Rathasamuth W, Tongsima S. Discovery of significant porcine SNPs for swine breed identification by a hybrid of information gain, genetic algorithm, and frequency feature selection technique. BMC Bioinformatics. 2020;21:1-28.
Pérez-Enciso M, Zingaretti LM. A guide on deep learning for complex trait genomic prediction. Genes. 2019;10(7):553.
Schiavo G, Bertolini F, Galimberti G, Bovo S, Dall’Olio S, Costa LN, Gallo M, Fontanesi L. A machine learning approach for the identification of population-informative markers from high-throughput genotyping data: application to several pig breeds. Animal. 2020;14(2):223-32.
Seo D, Lee DH, Choi N, Sudrajad P, Lee SH, Lee JH. Estimation of linkage disequilibrium and analysis of genetic diversity in Korean chicken lines. PLoS One. 2018;13(2):e0192063.
Shim JM, Seo DW, Seo S, Kim JJ, Min DM, Kim J, Jeon JT, Lee JH. Discrimination of Korean cattle (Hanwoo) with imported beef from USA based on the SNP markers. Korean J Food Sci Anim Resour. 2010;30:918-22.
Singh A, Thakur N, Sharma A. A review of supervised machine learning algorithms. 2016 3rd International Conference on Computing for Sustainable Global Development (INDIACom). 2016:1310-5. Ieee.
Team RC. R: A language and environment for statistical computing. 2013.
Tharwat A. Linear vs. quadratic discriminant analysis classifier: a tutorial. Int J Appl Pattern Recog. 2016;3(2):145-80.
Vignal A, Milan D, SanCristobal M, Eggen A. A review on SNP and other types of molecular markers and their use in animal genetics. Genet Sel Evol. 2002;34(3):275-305.
Yeung RM, Morris J. Consumer perception of food risk in chicken meat. Nutr Food Sci. 2001

Table 1. Information of samples used in this study.

Chicken Group	Population Code	Origin of population	Description	600k array	Validation
Government maintained Chicken (NIAS)	NC	Rhode Island Red	Imported (1960's) and adapted chicken population	6
	ND	Rhode Island Red		6
	NH	Cornish		6
	NS	Cornish		6
	NR	Redbrown Korean native chicken	Purebred of Korean native chicken	6
	NY	Yellowbrown Korean native chicken	Purebred of Korean native chicken	5
Commercial native chicken	HH	Hanhyup Farm	Founder population for new chicken breeding stock	23	36
	HF			23	36
	HY			21	26
	HW		Maintained population	23
	HS			23
	HG			23
	HV			23
	HA			20
	HZ			15
	1E		HH, HF, HY cross		10
	2C		HH, HF, HY cross		10
	WM_2	Woorimatdaq ver2	NIAS developed cross chicken population		10
	Yelim K	Yellim Farm	Private population		5
	HI	Hyunin Farm	Private population		5
Commercial Broiler	Ab	Arbor Acre	Meat type chicken	10	11
	Cobb	Cobb broiler		12	8
	Ross	Ross broiler		12	20
Commercial Layer	LO	Lohmann brown	Egg type chicken	10	5
Commercial Layer	HL	Hyline brown	Egg type chicken	10
Total				283	182

Table 2. The accuracy of identification of targeted (case) chicken population using selected minimum number of SNP markers (Feature Selection) by different machine learning models.

Classification Model	Accuracy	auc	Precision	Sensitivity	Specificity
Ada Boost 33 selected markers
Decision Tree	0.995		1	0.992	1
AdaBoost	0.995		1	0.992	1
Linear SVM	0.995		1	0.992	1
QDA	0.995		1	0.992	1
Random Forest	0.995		1	0.992	1
LDA	1		1	1	1
K-Nearest Neighbors	1		1	1	1
Naïve Bayes	0.995		1	0.992	1
Random Forest 44 selected markers
Decision Tree	0.969		1	0.949	1
AdaBoost	0.969		1	0.949	1
Linear SVM	0.990		1	0.983	1
QDA	0.995		1	0.992	1
Random Forest	0.969		1	0.949	1
LDA	1		1	1	1
K-Nearest Neighbors	0.984		1	0.975	1
Naïve Bayes	0.969		1	0.949	1
Decision Tree 8 selected markers
Decision Tree	0.984		1	0.975	1
AdaBoost	0.984		1	0.975	1
Linear SVM	0.964		1	0.941	1
QDA	0.974		1	0.958	1
Random Forest	0.979		0.991	0.975	0.986
LDA	0.995		1	0.992	1
K-Nearest Neighbors	0.984		1	0.975	1
Naïve Bayes	0.969		1	0.949	1

SVM: Support Vector Machine, QDA: Quadratic Discriminant Analysis, LDA: Linear Discriminant Analysis

Table 3. Prediction of validation samples using selected 36 SNP markers.

Pop	N	Ada Boost	Random Forest	Decision Tree	Linear Discriminant Analysis	Naïve Bayes	Nearest Neighbor	Quadratic Discriminant Analysis
HH	36	1	1	0.972	1	1	0.972	0.972
HF	36	1	1	0.972	1	1	1	1
HY	26	0.962	0.885	0.769	0.962	1	0.962	0.923
1E	10	1	0.7	0.9	1	1	1	1
2C	10	1	1	0.8	1	1	1	0.9
Ab	11	1	1	1	1	1	1	1
Cobb	8	1	1	1	1	1	1	1
Ross	20	1	1	1	1	1	1	1
LO	5	1	1	1	1	1	1	1
WM_2	10	1	1	1	1	1	1	1
Yelim K	5	1	1	1	1	1	1	1
HI	5	1	11	1	1	1	1	1

Pop: population, N: Number of predicted samples, HH: Hanhyup H, HF: Hanhyup F, HY: Hanhyup Y, 3way CC (HH, HF, HY): 1E and 2C, Ab: Abor Acre, Cobb: Cobb broiler, Ross: Ross broiler, LO: Lohmann brown, WM_2: Woorimatdaq ver2 commercial chicken, Yelim K: Yelim farm commercial chicken, HI: Hyunin commercial chicken

Download PDF

Version 1

posted

You are reading this latest preprint version

Identification of Chicken Populations by Machine Learning Models Using the Minimum Number of SNPs

Status:

Version 1

Abstract

Figures

Background

Methods

Results

Discussion

Conclusions

Abbreviations

Declarations

References

Tables

Supplementary Files

Status:

Version 1