Accuracy of genomic selection for alfalfa biomass yield in different reference populations

Abstract

Background: Genomic selection based on genotyping-by-sequencing (GBS) data could accelerate alfalfa yield gains, if it displayed moderate ability to predict parent breeding values. Its interest would be enhanced by predicting ability also for germplasm/reference populations other than those for which it was defined. Predicting accuracy may be influenced by statistical models, SNP calling procedures and missing data imputation strategies.

Results: Landrace and variety material from two genetically-contrasting reference populations, i.e., 124 elite genotypes adapted to the Po Valley (sub-continental climate; PV population) and 154 genotypes adapted to Mediterranean-climate environments (Me population), were genotyped by GBS and phenotyped in separate environments for dry matter yield of their dense-planted half-sib progenies. Both populations showed no sub-population genetic structure. Predictive accuracy was higher by joint rather than separate SNP calling for the two data sets, and using random forest imputation of missing data. Highest accuracy was obtained using Support Vector Regression (SVR) for PV, and Ridge Regression BLUP and SVR for Me germplasm. Bayesian methods (Bayes A, Bayes B and Bayesian Lasso) tended to be less accurate. Random Forest Regression was the least accurate model. Accuracy attained about 0.35 for Me in the range of 0.30-0.50 missing data, and 0.32 for PV at 0.50 missing data, using at least 10,000 SNP markers. Cross-population predictions based on a smaller subset of common SNPs implied a relative loss of accuracy of about 25% for Me and 30% for PV. Genome-wide association analyses based on large subsets of M. truncatula-aligned markers revealed many SNPs with modest association with yield, and some genome areas hosting putative QTLs. A comparison of genomic vs. conventional selection for parent breeding value assuming 1-year vs. 5-year selection cycles, respectively, indicated over three-fold greater predicted yield gain per unit time for genomic selection.

Conclusions: Genomic selection for alfalfa yield is promising, based on its moderate prediction accuracy, moderate value of cross-population predictions, and lack of sub-population structure. There is limited scope for searching individual QTLs with overwhelming effect on yield. Some of our results can contribute to better design of genomic selection experiments for alfalfa and other crops with similar mating systems.

Fig. 1

Number of SNP markers for different genotype missing data thresholds and SNP calling strategies. Results for 124 genotypes of the Po Valley population (PV) and 154 genotypes of the Mediterranean population (Me) subjected to separate SNP calling (data sets PV_Sep and Me_Sep), or joint SNP calling with subsequent application of missing data thresholds to separate populations (PV_Joint and Me_Joint) or joint populations (COMMON)

Fig. 2

STRUCTURE analysis of sub-populations. Log likelihood values of posterior probability as a function of the number of sub-populations, separately for the Po Valley (PV) and the Mediterranean (Me) populations

Fig. 3

Prediction accuracy for different genotype missing data imputation methods, SNP calling strategies and missing data thresholds. Results for four imputation methods (MNI, Mean imputation; SVDI, Singular value decomposition imputation; RFI, Random forest imputation; LHCI, Localized haplotype clustering imputation) applied to Po Valley (PV) and Mediterranean (Me) data sets subjected to separate SNP calling (PV_Sep and Me_Sep) or joint SNP calling (PV_Joint and Me_Joint), using Support Vector Regression with linear kernel

Fig. 4

Prediction accuracy of four genomic selection models at different genotype missing data thresholds. Results for Support Vector Regression with linear (SVR-lin) and gaussian (SVR-gau) kernel, Random Forest Regression (RFR), Ridge Regression BLUP (rrBLUP), Bayes A, Bayes B and Bayesian Lasso models applied to Po Valley (PV_Joint) and Mediterranean (Me_Joint) data sets subjected to joint SNP (random forest imputation of missing data)

Fig. 5

Accuracy of genomic selection for intra-population and cross-population prediction strategies at different genotype missing data thresholds. Intra-population prediction using all markers subjected to joint SNP calling (PV_Joint and Me_Joint data sets) or only markers satisfying the common filtering criteria (COMMON data set), and cross-population predictions using the COMMON data set, for Po Valley (PV) and Mediterranean (Me) populations, using Support Vector Regression with linear kernel or Ridge Regression BLUP (random forest imputation of missing data)

Fig. 6

Association (Manhattan plot) of M. truncatula-aligned SNP markers with total dry matter yield. Results for Po Valley (PV) and Mediterranean (Me) populations