Using unsupervised learning techniques to assess interactions among complex traits in soybeans

Soybean yield components and agronomic traits are connected through physiological pathways that impose tradeoffs through genetic and environmental constraints. Our primary aim is to assess the interdependence of soybean traits by using unsupervised machine learning techniques to divide phenotypic associations into environmental and genetic associations. This study was performed on large scale, jointly analyzing 14 quantitative traits in a large multi-parental population designed for genetic studies. We collected phenotypes from 2012 to 2015 from a soybean nested association panel with 40 families of approximately 140 individuals each. Pearson and Spearman correlations measured phenotypic associations. A multivariate mixed linear model provided genotypic and environmental correlations. To evaluate relationships among traits, the study used principal component and undirected graphical models from phenotypic, genotypic, and environmental correlation matrices. Results indicate that high phenotypic correlation occurs when traits display both genetic and environmental correlations. In genetic terms, length of reproductive period, node number, and canopy coverage play important roles in determining yield potential. Optimal grain yield production occurs when the growing environment favors faster canopy closure and extended reproductive length. Environmental associations found among yield components give insight into the nature of yield component compensation. The use of unsupervised learning methods provides a good framework for investigating interactions among various quantitative traits and defining target traits for breeding.