Assessment of Genomic Prediction Strategies after Animal Genome-Wide Association Study

doi:10.21203/rs.3.rs-2331918/v1

Download PDF

Research Article

Assessment of Genomic Prediction Strategies after Animal Genome-Wide Association Study

https://doi.org/10.21203/rs.3.rs-2331918/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

The detection of candidate variants with interesting traits is a major goal of a genome-wide association study (GWAS). GWAS-associated markers are considered candidate functional loci regarding animal and plant breeding and can serve to predict and treat human genetic diseases. Significant selected markers are functionally validated via molecular biology experiments or statistically validated by genomic prediction (GP) in an individual population. GWAS in a whole population used for GP causes an overprediction regarding accuracy. However, whether this overprediction exists in any traits with different genetic architectures remains unknown, while the extent of the difference between overprediction and actual prediction is also undetermined. The lack of whole key genetic information and linear dependence ubiquity can make perfect prediction of traits of interest impossible. A stable and adaptable prediction method for multiple genetic architectures is thus essential.

Results

We used a public dataset to present the accuracy bias in a cross-validation population with different genetic architectures and developed an approach termed “marker-assisted best linear unbiased prediction (MABLUP),” with removed linear dependence to improve the prediction accuracy for complex traits with genetic architectures. The MABLUP showed better prediction accuracy than other methods for traits under the control of few quantitative trait nucleotides (QTNs) and similar prediction accuracy to the best-known methods for traits under many QTNs.

Conclusions

The reasonable design of GP in the cross-validation after animal GWAS can be used to present actual potential breeding ability of detected significant markers. The MABLUP is a more stable and accurate GP method for more complex genetic traits.

genome-wide association study

genomic prediction

overprediction

complex genetic traits

MABLUP

Genome-wide association studies (GWAS) are an important tool for the detection of candidate variants associated with traits of interest [1, 2]. More than a million single nucleotide polymorphisms (SNPs) associated with production, performance, resistance, and genetic disease have been identified using GWAS [3–8]. In essence, GWAS is a process that filters markers of associated variants from whole genome data. These markers could potentially improve or develop animal and plant breeding or assist in prediction and treatment of human genetic diseases.

GWAS are not only used for associated marker detection but also for the selection of potential individuals by the genotype of their associated markers. However, these associated markers explain only a limited portion of heritability; this phenomenon was named “missing heritability” [9, 10]. The lack of key genetic information and ubiquity of linear dependence can make the perfect prediction of traits of interest impossible.

Another challenge is the validation of GWAS results. Although other powerful and efficient GWAS tools have been developed, including Bayesian-information and Linkage-disequilibrium (LD) iteratively nested keyway (BLINK), fixed and random model circulating probability unification (FarmCPU), and multiple loci mixed models (MLMM), the effectiveness and availability of GWAS results remain debatable. Molecular approaches for validating these significant markers include disturbing expression and gene knockout. However, these methods typically prove difficult for laboratories. Many studies have performed the validation of genome prediction (GP) using significant markers in another individual population [11–13].

In addition to integrating GWAS and GP, this study evaluated significant markers by GWAS and used these to predict genomic estimated breeding values (GEBVs). Cross-validation (CV) should be performed before GWAS to avoid overprediction regarding the use of phenotypes in the inference. Although biased estimation of prediction accuracy in the inference population have been reported [14, 15], it is unclear whether this overprediction exists for any traits with different genetic architectures. The extent of the difference between overprediction and actual maximum prediction accuracy is also unknown.

In this study, we reviewed the process of validating GWAS results in the GP and combined a marker-assisted best linear unbiased prediction (MABLUP) with linear removed markers selected by GWAS as the fixed effect and the kinship created by whole markers as the random effect. Our objectives were to (1) assess the differences between overprediction and actual maximum prediction accuracy following GWAS, (2) overcome the limitation of linear dependence among selected markers, and (3) combine these methods to provide wide applicability and accurate prediction using selected markers.

Accuracy Assessment of Different Validation Strategies

Accuracy assessment results are shown in Fig. 1. Following the increasing numbers of QTNs, all prediction accuracies were decreased. Following the decreasing heritability trend, all prediction accuracies were debased. For simulated traits belonging to simple Mendelian traits (using any of the validation strategies), the prediction accuracy remained similar. In any genetic background, “ACC with Reals” presented the maximum prediction accuracy. The “ACC without CV,” “ACC Holds,” and “ACC within CV” strategies are listed in decreasing order of prediction accuracy. In line with the increasing number of QTNs, the gaps among predicted accuracies showed a gradual increase. These gaps clearly demonstrated the impact of overprediction in the “ACC without CV” set. The “ACC Holds” set corresponds to the maximum prediction accuracy in the CV population before GWAS. This is due to the fact that the number of individuals in the reference population can reach the maximum value, which is equal to the number in the whole population minus one. In this type of clustered group—although there is only one individual difference between the reference and inference populations—the GWAS results were different between the reference population and the whole population. The “ACC within CV” set corresponded to the minimum prediction accuracy. Nevertheless, this strategy is not overpredicted and is a more efficient approach to determine whether significant markers from GWAS could be effectively used in breeding selection. “ACC within CV” running time was equal to 250 loops (50*5); however, the running times of “ACC Holds” were equal to 1939 loops (1940 − 1).

Using Significant Markers to Predict Kinship and Covariance

To compare the accuracy of predicted phenotypes between using significant markers as kinship and covariance, we simulated a trait with 0.5 heritability and 100 QTNs. The leave-one-out CV method was used to divide the whole population into 1940 groups, each including one individual. The all-predicted phenotype from 1940 loops is plotted in Fig. 2. The predicted phenotype values are shown against the observed phenotype values using kinship as a random effect in gBLUP. The Licols function approach achieved 60% accuracy for correlation between observed phenotypes and predicted phenotypes. The kinship approach in turn reached 33% accuracy. Evidently, the comparison between the fixed linear model (LM) (Licols function approach) and random LM (kinship approach) demonstrated the computing time of the former was considerably faster than that of the latter.

MABLUP Accuracy with Different Genetic Architectures

To fully understand the advantage of the MABLUP approach under different genetic architectures, we examined the predicted accuracy on simulated traits with heritability and number of QTNs. We combined MABLUP with multiple testing GWAS, including BLINK, MLMM, and FarmCPU; we also combined the LM with these three methods (Fig. 3). The predictive accuracies were used to compare rrBLUP with Bayesian LASSO. We additionally presented the accuracy difference between MABLUP and gBLUP in a small view of number of QTNs (Additional file 2). For any given situation of GWAS methods incorporating heritability and number of QTNs, the MABLUP approach was better than the LM approach. However, MABLUP combined with BLINK was superior to the other two GWAS methods. All predicted methods after GWAS were better than rrBLUP and Bayesian LASSO for simple traits controlled by a small number of QTNs. However, the MABLUP approach had a similar accuracy to rrBLUP and Bayesian LASSO for complex traits controlled by a large number of QTNs.

GWAS Before GS

GWAS result evaluation depends on several factors, including the effectiveness and availability of significant markers. First, sufficient population sizes are necessary to provide reliable and acceptable results. Next, the selected GWAS model and parameters can be used to evaluate whether the testing is suitable and logical. Finally, a major concern posed for genetic researchers is whether the detected significant markers can be useful for animal or plant breeding or human genetic disease treatment and prediction.

Several research efforts have validated significant markers from GWAS in an independent population [16–18]. Although this strategy makes the validated population independent from the testing GWAS population, it loses power in the testing GWAS step, because the whole population, including the validated and testing GWAS populations, may contain more effective individuals than any subpopulation. Another strategy is using whole genome markers after GWAS to predict phenotype values of the inference population under CV [17, 19–24]. Whole genome markers are difficult and expensive to use to evaluate breeding values or select a hybridization combination. However, the selected significant markers from GWAS have reportedly improved the predicted accuracy [11, 25], which implies that subset filtered markers could potentially improve breed planning better than whole genome information.

Marker-Assisted Selection

Unlike GP, marker-assisted selection (MAS) uses only a few markers to predict phenotype and breeding values [26, 27]. The markers used herein were selected from past studies, including GWAS, QTL mapping, and other molecular biology experiments [26, 28]. In our previous work, we identified a model with real QTNs as the most accurate prediction model [29]. Notably, the pseudo QTNs selected by GWAS exhibited an improved prediction accuracy, compared to a model with whole markers for simple traits. However, the use of pseudo QTNs selected by GWAS presents complications, in that linear dependence is widespread in the whole genome. Previously, we used markers selected by GWAS to build kinship and predict phenotype values; these were named “sBLUP.” The “sBLUP” markers were selected from bins containing neighboring markers in each window size. However, the linear dependent markers will always be present in a whole chromosome or genome. In the linear dependent removal step, the order of the effect of all selected markers should also be considered.

MAS with Kinship

For the most complex traits, MAS does not provide improved prediction compared to GP [26, 30, 31]. The genetic variance explained by MAS-selected markers is considerably lower than that by whole markers. This phenomenon is known as “missing heritability” [32–34]. The lack of any effective QTNs is a major factor in this phenomenon. Hence, the GWAS methods show reduced detection power for complex traits. GWAS and MAS are also beneficial for Mendelian traits. For complex traits, the kinship between individuals still showed fair prediction ability because complex traits are controlled by polygenes.

We combined MAS and gBLUP with linear dependence removed, significant markers as fixed effects, and whole genome markers as random effects and named this approach “MABLUP.” This method was first reported by Lopes et al. [35] and was tested with a pig dataset, including four populations and a real phenotype, namely, number of teats. Herein, we compared the MABLUP approach with the LM approach, rrBLUP, and Bayesian LASSO for multiple simulated traits under different genetic backgrounds. When simple traits were tested, GWAS was useful and powerful; therefore, the accuracy of MABLUP was better than those of rrBLUP and Bayesian LASSO. However, when complex traits were tested, the kinship with whole genome markers effectively explained the phenotype, implying that MABLUP accuracy is similar to that of rrBLUP. Regarding applications, MABLUP should be preferably considered for molecular breeding planning.

The validation of significant markers after GWAS is considerably more effective in MAS than in GP. CV before GWAS helps to avoid overprediction. Although GWAS was used for single-marker testing or multiple-loci testing, MABLUP, with linearly dependent markers removed, presented considerable advantages and applicability compared to the linear prediction model, such as rrBLUP, and Bayesian LASSO methods.

Genotype Data

Murine data were obtained from a heterogeneous stock population kept by The Welcome Trust Centre for Human Genetics (available at http://mtweb.cs.ucl.ac.uk/mus/www/mouse/HS/index.shtml). This population was generated by crossing eight inbred lines, followed by 50 generations of random mating. The dataset contained 1940 individuals genotyped with 12,227 SNP markers [36]. The average interval between the SNPs was 167 kilobase-pairs (kb) (SD = 141.5), where 99.0% and 81.2% of the intervals were below 500 kb and below 250 kb, respectively.

Simulated Phenotype

Simulated assessment of the prediction strategy requires a large set of phenotypes of different genetic architectures. Based on presupposed genetic parameters, we used the GAPIT “Phenotype.Simulation” function to create simulation traits [37]. The heritability parameter was termed “h2,” and the quantitative trait nucleotide (QTN) number parameter was termed “NQTN.” The total summation of marker effects was defined as GEBV of an individual [38]. The ratio between random residual values in the whole population and individual GEBVs was defined as presupposed heritability. All simulated experiments were repeated 50 times to minimize bias by random selection.

Validation and Overpredicting

To validate overprediction and determine the difference between various genetic architectures, we designed four prediction strategies (Additional file 1). The first strategy, termed “accuracy with real QTNs (ACC with Reals),” kept real QTNs separate from simulation traits to predict inference phenotype values under 5-fold CV. The second strategy, named “accuracy with holding CV (ACC Holds),” used leave-one-out CV to divide the whole population into 1940 groups, each including one individual, and detected the association markers in the reference population (1939 individuals) [39, 40]. The significant markers were then used to predict inference phenotype (one individual only). After 1940 iterations (each individual was predicted once), whole phenotype values and predicted values were used to calculate the Pearson correlation, representing accuracy. The third strategy, “accuracy within CV (ACC within CV),” used 5-fold CV to divide the whole population into five groups and detected the association markers in the reference population (four groups). The average correlations in all folds between phenotype values and prediction values were considered accurate. The last strategy, “ACC without CV,” involved GWAS in the whole population and detected significant SNPs. Following the same clustered group from the 5-fold CV, the total significant marker genotype matrix and the phenotype values from reference populations were used to predict the phenotype values of the inference population. All simulated experiments were randomly repeated 50 times to reduce noise due to random selection.

All significant marker genotypes from the second and third strategies were used to create kinship to predict inference phenotype values. The correlation between the real phenotype and predicted phenotype was considered as accuracy. The average values of accuracy in repeats and folds were used to define the final accuracy. The heritability gradient was set at 0.25, 0.5, and 0.75 for low, median, and high heritability, respectively. QTN number gradients were set at 20, 100, and 200 for simple, polygene, and complex traits, respectively.

Markers in Linear Dependence

Based on the linkage disequilibrium and population stratification, there was considerable number of linearly dependent markers in the whole genome [41, 42]. Following GWAS, linearly dependent markers remained in the significant loci. Their genotype matrix cannot be solved using an inverse matrix as a full-rank matrix. This would allow significant markers to be directly added into kinship to predict phenotype in the mixed LM (MLM); however, this method would attribute the lack of prediction accuracy to the plethoric relative variables. Herein, an approach to filter out the markers with un-full rank in the whole markers genotype matrix was coded. First, the matrix with colinear markers genotype was decomposed into an orthogonal matrix (Q) and triangular matrix ® using QR decomposition. The positions of diagonal non-zero elements of the R matrix were used to mark as non-colinear markers. Subsequently, the positions of diagonal zero elements of the R matrix were used to mark as colinear markers. This filtering method was coded into a GAPIT sub-function, termed “GAPIT.Licols.”

Next, we compared the prediction ability between this filtering function and the kinship with all significant markers in a simulated trait, which simulated heritability < 0.5 and 100 QTNs. The leave-one-out CV method was used to perform prediction, and the MLM was used to perform GWAS. The top three principal components were used for the covariance matrix in the GWAS for the reference population, each repeat including 1939 individuals. The predicted model was termed “genomic best linear unbiased prediction (gBLUP).”

MABLUP

Based on the Licols function in the GAPIT function, the significant markers were listed as a fixed effect of the MLM. However, GWAS can only detect potential markers in a situation of relatively strong heritability, without large numbers of causal markers. To obtain wide adaptability and practicability of the GP method, the relationship matrix between individuals was used to estimate overall GEBVs. The significant markers (filtered by Licols) were integrated as a fixed effect and the kinship from all markers as a random effect into the MLM. This method was named “marker-assisted BLUP (MABLUP).”

To compare the prediction accuracy between the MABLUP and LM strategies, three multiple testing GWAS methods, including Bayesian-information and BLINK, FarmCPU, and an MLMM, were used to detect association markers. These three testing models showed considerably greater power than other GWAS methods. All significant markers from the GWAS methods were used to predict GEBVs using the MABLUP and LM strategies. In contrast, the GP methods, rrBLUP and Bayesian LASSO, were used to compare predictive ability.

BLINK, Bayesian-information and Linkage-disequilibrium iteratively nested keyway; CV, cross-validation; FarmCPU, fixed and random model circulating probability unification; GEBV, genomic estimated breeding value; GP, genome prediction; GWAS, genome-wide association studies; LD, linkage-disequilibrium; LM, linear model; MABLUP, marker-assisted best linear unbiased prediction; MAS, marker-assisted selection; MLM, mixed linear model; MLMM, multiple loci mixed models; PC, principal component; SNP, single nucleotide polymorphism; QTN, quantitative trait nucleotide

Ethics approval and consent to participate

Not applicable

Consent for publication

Not applicable

Availability of data and materials

The MABLUP code is available in the GAPIT github repository (https://github.com/jiabowang/GAPIT3).

Competing interests

The authors declare that they have no competing interests.

Funding

This project was partially funded by the Qinghai Science and Technology Program, China (2022-NK-110), Sichuan Science and Technology Program, China (Award #s 2021YJ0269 and 2021YJ0266), the Fundamental Research Funds for the Central Universities, China (Southwest Minzu University, ZYN2022023), and the Program of Chinese National Beef Cattle and Yak Industrial Technology System, China (CARS-37).

Authors’ contributions

JW: software, data curation, writing first draft of manuscript, visualization, testing, methodology, supervision, and validation. YK, LC, and WP: Writing first draft of manuscript. JZ: Conceptualization and Manuscript Revision.

Acknowledgements

We would like to thank Editage (www.editage.cn) for English language editing.

Hickey JM. Genome-Wide Association Studies and Genomic Prediction. Gondro C, van der Werf J, Hayes B, editors. Totowa: Humana Press; 2013;p. 1019.
Flint J, Eskin E. Genome-wide association studies in mice. Nat Rev Genet Nature Publishing Group. 2012;13:807–17.
Fang C, Ma Y, Wu S, Liu Z, Wang Z, Yang R, et al. Genome-wide association studies dissect the genetic networks underlying agronomical traits in soybean. Genome Biol. Genome Biology; 2017;18.
Tsai HY, Hamilton A, Tinch AE, Guy DR, Gharbi K, Stear MJ, et al. Genome wide association and genomic prediction for growth traits in juvenile farmed Atlantic salmon using a high density SNP array. BMC Genomics. BMC Genomics; 2015;1–9.
Cao X, Cheng J, Huang Y, Lan X, Lei C, Chen H. Comparative Enhancer Map of Cattle Muscle Genome Annotated by ATAC-Seq. Front Vet Sci. 2021;8:1–12.
Huang X, Wei X, Sang T, Zhao Q, Feng Q, Zhao Y, et al. Genome-wide association studies of 14 agronomic traits in rice landraces. Nat Genet. 2010/10/26. 2010;42:961–7.
Xiongwei L, Jiabo W, Mingshen S, Jingyi Z, Minghao Z, Jihong D, et al. Single Nucleotide Polymorphism Detection for Peach Gummosis Disease Resistance by Genome-Wide Association Study. Front Plant Sci. 2022;07.
Zhao M, Liu S, Pei Y, Jiang X, Jaqueth JS, Li B, et al. Identification of genetic loci associated with rough dwarf disease resistance in maize by integrating GWAS and linkage mapping. Plant Sci Elsevier B V. 2022;315:111100.
Connolly S, Heron EA. Review of statistical methodologies for the detection of parent-of-origin effects in family trio genome-wide association data with binary disease traits. Brief Bioinform. 2014;16:429–48.
Yang J, Lee T, Kim J, Cho MC, Han BG, Lee JY, et al. Ubiquitous Polygenicity of Human Complex Traits: Genome-Wide Analysis of 49 Traits in Koreans. PLoS Genet. 2013;9.
Shi A, Bhattarai G, Xiong H, Avila CA, Feng C, Liu B, et al. Genome-wide association study and genomic prediction of white rust resistance in USDA GRIN spinach germplasm. Hortic Res. 2022;9:1–17.
Srivastava RK, Singh RB, Pujarula VL, Bollam S, Pusuluri M, Chellapilla TS, et al. Genome-Wide Association Studies and Genomic Selection in Pearl Millet: Advances and Prospects. Front Genet. 2020;10:1–10.
Ballesta P, Bush D, Silva FF, Mora F. Genomic predictions using low-density SNP markers, pedigree and GWAS information: A case study with the non-model species Eucalyptus cladocalyx. Plants. 2020;9.
Runcie D, Cheng H. Pitfalls and remedies for cross validation with multi-trait genomic prediction methods. G3 Genes, Genomes. Genet. 2019;9:3727–41.
McGowan M, Wang J, Dong H, Liu X, Jia Y, Wang X, et al. Ideas in Genomic Selection with the Potential to Transform Plant Molecular Breeding. Plant Breed Rev. 2021. p. 273–319.
Slavney AJ, Kawakami T, Jensen MK, Nelson TC, Sams AJ, Boyko AR. Five genetic variants explain over 70% of hair coat pheomelanin intensity variation in purebred and mixed breed domestic dogs. PLoS ONE. 2021;16:1–23.
Yu X, Li X, Guo T, Zhu C, Wu Y, Mitchell SE, et al. Genomic prediction contributing to a promising global strategy to turbocharge gene banks. Nat Plants. 2016;2:16150.
Kiser JN, Keuter EM, Seabury CM, Neupane M, Moraes JGN, Dalton J, et al. Validation of 46 loci associated with female fertility traits in cattle. BMC Genomics. BMC Genomics; 2019;20.
Zia B, Shi A, Olaoye D, Xiong H, Ravelombola W, Gepts P, et al. Genome-Wide Association Study and Genomic Prediction for Bacterial Wilt Resistance in Common Bean (Phaseolus vulgaris) Core Collection. Front Genet. 2022;13:1–13.
Ober U, Ayroles JF, Stone EA, Richards S, Zhu D, Gibbs RA, et al. Using Whole-Genome Sequence Data to Predict Quantitative Trait Phenotypes in Drosophila melanogaster. PLoS Genet. 2012/05/10. 2012;8:e1002685.
Daetwyler HD, Calus MPL, Pong-Wong R, de los Campos G, Hickey JM. Genomic prediction in animals and plants: Simulation of data, validation, reporting, and benchmarking. Genetics. 2013;193:347–65.
González-Recio O, Rosa GJM, Gianola D. Machine learning methods and predictive ability metrics for genome-wide prediction of complex traits. Livest Sci. 2014;166:217–31.
Chiaia HLJ, Peripoli E, Silva RM, de O, Aboujaoude, Feitosa C, FLB, Lemos MVA de, et al. Genomic prediction for beef fatty acid profile in Nellore cattle. Meat Sci Journal of Applied Genetics. 2017;128:60–7.
Ge F, Jia C, Bao P, Wu X, Liang C, Yan P. Accuracies of genomic prediction for growth traits at weaning and yearling ages in yak. Animals. 2020;10:1–11.
Yoshida GM, Yáñez JM. Increased accuracy of genomic predictions for growth under chronic thermal stress in rainbow trout by prioritizing variants from GWAS using imputed sequence data. Evol Appl. 2022;15:537–52.
Zhang Z, Zhang Q, Ding XD. Advances in genomic selection in domestic animals. Chin Sci Bull. 2011;56:2655–63.
Heffner EL, Lorenz AJ, Jannink JL, Sorrells ME. Plant breeding with Genomic selection: Gain per unit time and cost. Crop Sci. 2010;50:1681–90.
Andreescu C, Avendano S, Brown SR, Hassen A, Lamont SJ, Dekkers JCM. Linkage disequilibrium in related breeding lines of chickens. Genetics. 2007;177:2161–9.
Wang J, Zhou Z, Zhang Z, Li H, Liu D, Zhang Q, et al. Expanding the BLUP alphabet for genomic prediction adaptable to the genetic architectures of complex traits. Heredity (Edinb) Springer US. 2018;121:648–62.
Heffner EL, Lorenz AJ, Jannink JL, Sorrells ME. Plant breeding with Genomic selection: Gain per unit time and cost. Crop Sci. 2010;50:1681–90.
Hayes B, Goddard M. Genome-wide association and genomic selection in animal breeding. Genome 2010/11/16. 2010;53:876–83.
Trait GC, Gcta A, Yang J, Lee SH, Goddard ME, Visscher PM. Genome-Wide Association Studies and Genomic Prediction. In: Gondro C, van der Werf J, Hayes B, editors. Methods Mol Biol. Totowa: Humana Press; 2013. pp. 215–36.
Vinkhuyzen AAE, Pedersen NL, Yang J, Lee SH, Magnusson PKE, Iacono WG, et al. Common SNPs explain some of the variation in the personality dimensions of neuroticism and extraversion. Transl. Psychiatry. 2012. p. e125.
Wainschtein P, Jain D, Zheng Z, Working TA. Assessing the contribution of rare variants to complex trait heritability from whole genome sequence data. Nat Genet. 2022.
Lopes MS, Bovenhuis H, van Son M, Nordbø Ø, Grindflek EH, Knol EF, et al. Using markers with large effect in genetic and genomic predictions. J Anim Sci. 2017;95:59–71.
Neves HHR, Carvalheiro R, Queiroz S. a. A comparison of statistical methods for genomic selection in a mice population. BMC Genet BMC Genetics. 2012;13:100.
Wang J, Zhang Z. GAPIT Version 3: Boosting Power and Accuracy for Genomic Association and Prediction. Genomics Proteomics Bioinformatics. 19: Beijing Institute of Genomics, Chinese Academy of Sciences and Genetics Society of China; 2021. pp. 629–40.
Vallejo RL, Silva RMO, Evenhuis JP, Gao G, Liu S, Parsons JE, et al. Accurate genomic predictions for BCWD resistance in rainbow trout are achieved using low-density SNP panels: Evidence that long-range LD is a major contributing factor. J Anim Breed Genet. 2018;135:263–74.
Lopes FB, Wu XL, Li H, Xu J, Perkins T, Genho J, et al. Improving accuracy of genomic prediction in Brangus cattle by adding animals with imputed low-density SNP genotypes. J Anim Breed Genet. 2018;135:14–27.
Zhou Y, Isabel Vales M, Wang A, Zhang Z. Systematic bias of correlation coefficient may explain negative accuracy of genomic prediction. Briefings Bioinforma. 2016.
Cardon LR, Bell JI. Association study designs for complex diseases. Nat Rev Genet. 2001;2:91–9.
Jannink JL, Lorenz AJ, Iwata H. Genomic selection in plant breeding: from theory to practice. Brief Funct Genomics. 2010;9:166–77.

No competing interests reported.

Supplementary.docx
Additional file 1 Figure S1 Format: PDF Title: The pipeline of different prediction strategies from same dataset. Description: All experiments based on the same dataset and randomly simulated 50 times. In the blue approach, the real quantitative trait nucleotides (QTNs) were kept after simulation. Based on same random group under 5-folds, the phenotypes and QTNs genotypes of four reference groups and the QTNs genotypes of last inference group were used to predict the phenotypes of inference. In the red and green approaches, the individuals were selected into reference and inference groups before GWAS. The significant SNPs genotype from whole population and the phenotype from reference population were used to predict inference phenotype. The difference between red and green approaches is the Cross-validation, one is leave-one-out, the other is 5-folds. In the yellow approach, the whole population were used to detect significant markers in GWAS. Then the whole population was divided into 5 groups same as 5-folds cross validation. The correlation between real phenotype and predicted phenotype was considered as accuracy for each approach. Additional file 2 Figure S2 Format: PDF Title: The accuracy of markers assistant prediction strategies between gBLUP and MABLUP. Description: The MABLUP (darkblue) was performed by BLINK GWAS and MABLUP Genomic Prediction strategy. All significant markers were filtered by linear dependent removing then set as fixed effect. The whole markers were created kinship matrix to reveal individuals’ relationship that is same as gBLUP (darkred). All experiments were performed using simulation traits under heritability (0.25, 0.5, and 0.75) and NQTN (30, 50, 70, and 90) with 50 times repeats and 5-folds cross validation.

Download PDF

Version 1

posted

You are reading this latest preprint version

Assessment of Genomic Prediction Strategies after Animal Genome-Wide Association Study

Status:

Version 1

Abstract

Figures

Background

Results

Accuracy Assessment of Different Validation Strategies

Using Significant Markers to Predict Kinship and Covariance

MABLUP Accuracy with Different Genetic Architectures

Discussion

GWAS Before GS

Marker-Assisted Selection

MAS with Kinship

Conclusions

Methods

Genotype Data

Simulated Phenotype

Validation and Overpredicting

Markers in Linear Dependence

MABLUP

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1