Assessing genomic selection models for soybean growth traits at vegetative stage under low Phosphorus availability

Phosphorus (P) deficiency is an important constraint to soybean yield, since it is a macronutrient that plays a substantial role in photosynthesis and plant growth.  Developing materials that are efficient for P acquisition is a way of boosting crop productivity in environments with such abiotic stress.  Genomic Selection has emerged as a promising tool in breeding approaches to assess complex traits related to nutrient acquisition and enable more precise and accurate selection. Thus, we aimed to compare different genomic selection models to predict traits evaluated during vegetative stage in a panel of 265 soybean accessions. Root (RB) and Shoot (SB) biomasses (g) were measured, as well as Quantum Efficiency of Photosystem II (ΦPSII), in a greenhouse at Iowa State University, Agronomy department, Iowa State - USA. The experiment was conducted in two trials of four weeks between April and June 2024. In each trial, the soybeans were grown in two hydroponic tubs with Half Strength Hoagland Solution containing low Phosphorus level (90% deficiency). 42K SNPs were used from the public available SoySNP50K data from the genotyped PI accessions. We fitted six genomic prediction models for each trait: kernel Hilbert spaces regressions (RKHS), Bayesian Ridge Regression (BRR), BayesA, BayesB, BayesC and Bayesian Least Absolute Shrinkage and Selection Operator (LASSO). Cross Validation was used with testing populations of 65 accessions to validate the prediction accuracy, repeating partition 10 times and adopting 12000 iterations. For all traits, models BayesA and BayesB showed the lowest values of Residual Variance (RB: 0.0053 and   0.0056; SB: 0.071 and 0.078; ΦPSII: 0.0046 and 0.005, respectively) and Deviance Information Criterion (DIC) (RB: -533 and   -529; SB: 152 and 156, respectively). LASSO showed low residual variance and DIC for RB (0.0055 and -528, respectively) and for ΦPSII (0.0049 and -580, respectively), but the highest values for SB (0.095 and 167, respectively). Predictive abilities were small to ΦPSII (0.10 to 0.14) and moderate to RB (0.4 to 0.43) and SB (0.41 to 0.42). RKHS, BRR, BayesC and LASSO had the highest accuracies, that didn’t differ statistically, while BayesA was the least accurate for all traits, and BayesB also had smaller values for RB (0.42) and ΦPSII (0.12). Overall, all models had similar performance for the evaluated parameters, except BayesA and BayesB for prediction accuracy. Considering higher predictive ability and lower Residual Variance and DIC, LASSO was more efficient to predict RB and ΦPSII, while both RKHS and BRR had the best combinations of parameters to select for SB under low P availability.