The causal meaning of genomic predictors and how it affects the construction and comparison of genome-enabled selection models

The additive genetic effect is arguably the most important quantity inferred in animal and plant breeding analyses. The term effect indicates that it represents causal information, which is different from standard statistical concepts as regression coefficient and association. The process of inferring causal information is also different from standard statistical learning, as the former requires causal (i.e. non-statistical) assumptions and involves extra complexities. Remarkably, the task of inferring genetic effects is largely seen as a standard regression/prediction problem, contradicting its label. This widely accepted analysis approach is by itself insufficient for causal learning, suggesting that causality is not the point for selection. Given this incongruence, it is important to verify if genomic predictors need to represent causal effects to be relevant for selection decisions, especially because applying regression studies to answer causal questions may lead to wrong conclusions. The answer to this question defines if genomic selection models should be constructed aiming maximum genomic predictive ability or aiming identifiability of genetic causal effects. Here, we demonstrate that selection relies on a causal effect from genotype to phenotype, and that genomic predictors are only useful for selection if they distinguish such effect from other sources of association. Conversely, genomic predictors capturing non-causal signals provide information that is less relevant for selection regardless of the resulting predictive ability. Focusing on covariate choice decision, simulated examples are used to show that predictive ability, which is the criterion normally used to compare models, may not indicate the quality of genomic predictors for selection. Additionally, we propose using alternative criteria to construct models aiming for the identification of the genetic causal effects.