Abstract
Modulating days to flowering is a key mechanism in plants for adapting to new environments, and variation in days to flowering drives population structure by limiting mating. To elucidate the genetic architecture of flowering across maize, a quantitative trait, we mapped flowering in five global populations, a diversity panel (Ames) and four half-sib mapping designs, Chinese (CNNAM), US (USNAM), and European Dent (EUNAM-Dent) and Flint (EUNAM-Flint). Using whole-genome projected SNPs, we tested for joint association using GWAS, resampling GWAS and two regional approaches; Regional Heritability Mapping (RHM) (1, 2) and a novel method, Boosted Regional Heritability Mapping (BRHM). Direct overlap in significant regions detected between populations and flowering candidate genes was limited, but whole-genome cross-population predictive abilities were ≤0.78. Poor predictive ability correlated with increased population differentiation (r = 0.41), unless the parents were broadly sampled from across the North American temperate-tropical germplasm gradient; uncorrected GWAS results from populations with broadly sampled parents were well predicted by temperate-tropical FSTs in machine learning. Machine learning between GWAS results also suggested shared architecture between the American panels and, more distantly, the European panels, but not the Chinese panel. Machine learning approaches can reconcile non-linear relationships, but the combined predictive ability of all of the populations did not significantly enhance prediction of physiological candidates. While the North American-European temperate adaption is well studied, this study suggest independent temperate adaptation evolved in the Chinese panel, most likely in China after 1500, a finding supported by differential gene ontology term enrichment between populations.
Footnotes
Kelly Swarts wrote the manuscript and led all analyses, performing those not otherwise indicated, and projected Hapmap 3.21 SNPs onto the GBS samples. Peter Bradbury contributed substantially to the development of BRHM, ran the RMIP analysis, projected the EUNAM populations and contributed intellectually to the discussion and edited the manuscript. Jeff Glaubitz supplied database information for the samples and GBS genotyping and curation for some of the projected datasets. Tiffany Ho ran preliminary machine learning analyses, Zachary Miller ran the Random Forest Classifier analyses and Lynn Johnson set up and curated the database for the machine learning predictors. Yongxiang Li, Yu Li, Tianyu Wang and Zhiwu Zhang contributed the CNNAM datasets and phenotypes. Eva Bauer and Chris-Carolin Schön contributed the EUNAM datasets. Cinta Romay and Edward Buckler contributed the Ames dataset. Cinta Romay additionally added the ZCN genes to the Dong et al flowering candidate list information and selected lines for the temperate and tropical Fst sets and Edward Buckler contributed the USNAM dataset and a substantial intellectual input to the BRHM method, machine learning, interpretations, and manuscript edits.