Quantifying the relationship between adaptive traits and agro-climatic conditions

Abstract

Durum wheat is an economically important and regularly eaten food for billions of people in the world. Consequently, wheat breeders over the past century have increased the productivity and adaptability via strong selection applied to genes controlling agronomical important traits. In the International Center for Agriculture Research in Dry Areas (ICARDA), genbanks are using Focused Identification of the Germplasm Strategy (FIGS) to find out and quantify relationships between agro-climatic conditions and the presence of specific traits. Hence, the study is aimed to investigate the predictive value of various types of long-term agro-climatic variables on the future values of different traits as well as the association between these traits and those of the different agro-climatic characteristics. Ordinary multiple linear regression with stepwise variable selection method, and multiple linear regression models with predictors selected by penalized methods with mean square error cross-validation as a model selection criterion, are used to analyze 238 durum wheat landraces which were chosen from the International Center for Agriculture Research in the Dry Areas (ICARDA) genebanks. Each of the models are fitted on Days to Heading, Days to Maturity, Plant Height, Grain Weight and Thousand Kernel Weight response variables with 57 predictor variables, independently. The penalized based models used data splitting into training on which the model is fitted and test data set on which the fitted model is validated. Ordinary least square and weighted least square estimation methods are also used for parameter estimation and prediction of post model selection. Findings implied that there is high multicollinearity among the predictor variables. It is found that there are some predictors which affect positively and some others affect negatively for Days to Heading, Days to Maturity, Plant Height and Grain Weight using both ordinary and shrinkage based models. Longitude affects significantly the Thousand Kernel Weight using the ordinary MLR model, However, there is no significant predictor which affects the Thousand Kernel Weight from the shrinkage based MLR models. But longitude affects it significantly using the ordinary MLR model. It is revealed that model with predictors selected by Elastic net method seem to have good prediction on the Plant Height for both OLS and WLS estimation methods, while the prediction from the Lasso based model is not that much reasonable. Furthermore, for the Days to Heading and Grain Weight showed that there seems better prediction as their predicted values increase continuously as a function of the actual values though there is considerable variability. However, the Lasso based model used for Thousand Kernel Weight is not predicting well. In conclusion, inferences and predictions by the ordinary MLR models are not trusted due to the presence of multicollinearity in the model fitting, and violation of some model assumptions after model fitting. However, predictions using the models with predictors selected by the shrinkage methods may be better as the effects of the variability on these methods are minimal. Moreover, the WLS methods might give more sensible predictions than the OLS estimation methods. Better predictions were found on the Plant Height, Days to Heading and grain Weight.