SIKIRU ADENIYI ATANDA
DEVELOPMENT OF EFFICIENT PROTOCOLS FOR TRAINING POPULATION DESIGN AND RESOURCE ALLOCATION FOR GENOMIC SELECTION IN CIMMYT TROPICAL MAIZE BREEDING PROGRAMME
ABSTRACT
The strategy for implementation of genomic selection at International Maize and Wheat Improvement Center (CIMMYT) Global maize breeding programme has been to calibrate models using information from full-sibs in a “test-half-predict-half” approach. Though effective, this approach has limitations, as it requires large full-sib populations and limits the ability to shorten variety testing and breeding cycle times. The objectives of this study were to (i) determine strategy enabling integration of historical data in CIMMYT genomic selection pipeline to accelerate early yield testing stage, (ii) determine optimal resource allocation in early implementation of genomic selection in a breeding programme, and (iii) determine whether family relationship can be used to achieve stable and improved prediction accuracy when spreading the cost of large training setsacross families. This study evaluated training set design to determine the most effective use of phenotypic records collected on relatives for genomic prediction using datasets containing 849 (DS1) and 1389 (DS2) doubled haploid lines evaluated as testcrosses in CIMMYT Kenya experimental stations; Kiboko and Kakamega during the rainy season (June to September) and Kiboko under managed drought during the off season (December to March) in 2017 and 2018, respectively. The efficiency of different genetic optimization criteria in selecting individuals from historical data with predictive ability of the new population were evaluated. The genetic optimization criteria evaluated in this study included: 1) mean of generalized coefficient of determination (CDmean), 2) effective chromosome segment segregating across populations (Me), 3) deterministic prediction accuracy (DPAH), 4) deterministic prediction accuracy where the heritability of the trait was substituted by reliability (DPAR), and 5) average genomic relationship between an individual of historical data and all individuals of a specific new population (Avg_GRM). The results in this study showed that there is merit in the use of diverse populations as training sets when optimized using algorithms to select individuals that were genetically related to prediction sets. In a breeding programme with limited historical records, phenotyping cost can be spread across close relatives by phenotyping only a small number of lines from each related population. This improves prediction accuracy compared to within bi-parental population prediction, especially when the training set for within full-sib prediction is small. Further, the sparse testing genomic selection strategy in which the genetic merit of new lines is evaluated in different but genetically correlated environments improved prediction accuracy compared to testhalf-predict-half genomic selection strategy and given that, all new lines have phenotypic data, it was seemingly robust in developing historical dataset. Finally, this study demonstrated that prediction accuracy in either sparse testing or “test-half-predict-half” can further be improved by using the CDmean genetic optimization algorithm to select individuals that maximize the genetic space of the population for phenotyping and which lines are to be only genotyped for advancement based on genomic prediction.