15Genetic architecture reflects the pattern of effects and interaction of genes underling 16 phenotypic variation. Most mapping and breeding approaches generally consider the additive 17 part of variation but offer limited knowledge on the benefits of epistasis which explains in 18 part the variation observed in traits. In this study, the cowpea multiparent advanced 19 generation inter-cross (MAGIC) population was used to characterize the epistatic genetic 20 architecture of flowering time, maturity, and seed size. In addition, considerations for 21 epistatic genetic architecture in genomic-enabled breeding (GEB) was investigated using 22 parametric, semi-parametric, and non-parametric genomic selection (GS) models. Our results 23 showed that large and moderate effect sized two-way epistatic interactions underlie the traits 24 examined. Flowering time QTL colocalized with cowpea putative orthologs of Arabidopsis 25 thaliana and Glycine max genes like PHYTOCLOCK1 (PCL1 [Vigun11g157600]) and 26 PHYTOCHROME A (PHY A [Vigun01g205500]). Flowering time adaptation to long and 27 short photoperiod was found to be controlled by distinct and common main and epistatic loci. 28Parametric and semi-parametric GS models outperformed non-parametric GS model. Using 29 known QTL as fixed effects in GS models improved prediction accuracy when traits were 30 controlled by both large and moderate effect QTL. In general, our study demonstrated that 31 prior understanding the genetic architecture of a trait can help make informed decisions in 32 GEB. This is the first report to characterize epistasis and provide insights into the 33 underpinnings of GS versus marker assisted selection in cowpea. 34
35Genomic selection, flowering time, and photoperiod. 37 reflect the entire variation responsible for the trait and may not be transferable to other 64 genetic backgrounds . Third, linkage mapping is limited in power to detect 65 small effect loci, thus only the available large effects loci are used for MAS (Ben-Ari and 66Lavi 2012). Notably, MAS is more efficient with traits controlled by few genomic loci and 67 not polygenic traits (Bernardo, 2008). In contrast, genomic selection (GS) that employs 68 genome wide markers has been found to be more suited for complex traits, and also having 69 higher response to selection than MAS (Bernardo and Yu, 2007;Wong and Bernardo, 2008; 70 Cerrudo et al., 2018). 71 4 In GS, a set of genotyped and phenotyped individuals are first used to train a model 72 that estimates the genomic estimated breeding values (GEBVs) of un-phenotyped but 73 genotyped individuals (Jannink, Lorenz and Iwata, 2010). GS models often vary in 74 performance with the genetic architecture of traits. Parametric GS models are known to 75 capture additive genetic effects but not efficient with epistatic effects due to the 76 computational burden of high-order interactions (Moore and Williams, 2009; Howard, 77 Carriquiry and Beavis, 2014). Parametric GS models with incorporated kernels (marker based 78 relationship matrix) f...