Evaluating the Interpretability of SNP Effect Size Classes in Bayesian Genomic Prediction Models

Fanny Mollandin (1) , Andrea Rau (1) , Pascal Croiseau (1)

1 GABI - Génétique Animale et Biologie Intégrative

Genomic prediction models are widely used as an evaluation tool for genomic selection in plant and animal breeding, and for the calculation of polygenic risk scores for human diseases. Among other approaches, non-linear Bayesian models represent an attractive approach to perform genomic prediction, due in part to their flexibility and ability to perform variable selection. In recent years, a suite of models in the so-called Bayesian alphabet have been proposed to find a compromise between reality, which may correspond to an omnigenic model for complex traits, and computational convenience. In particular, the BayesR model (Erbe et al., 2012) strikes this balance by modeling single nucleotide polymorphisms (SNPs) as a mixture of markers with null, small, medium, or large variance. Although these approaches have been shown to achieve improved prediction accuracy in a variety of scenarios, there is still a need to evaluate the extent to which the assignment of SNPs to specific effect size classes (small, medium, large) reflects the true underlying genetic architecture and is meaningful for downstream SNP selection. Based on a set of real genotypes, we generated simulated data under a wide variety of genetic architectures and phenotype heritabilities, and we evaluated the impact of excluding or including causal markers among the genotypes. We define several statistical criteria for QTL mapping, including several based on sliding windows to account for linkage disequilibrium, and we evaluated their ability to accurately prioritize known causal markers in various scenarios. Finally, we provide some insight into how the use of additional prior biological information can contribute to the interpretability of the SNP classes identified by BayesR. In particular, BayesRC (MacLeod et al., 2016), an extension of BayesR, categorizes SNPs according to their functional annotations, each of which is independently modeled using BayesR to allow for a varying enrichment of QTLs. The use and extension of the BayesRC model in conjunction with appropriate QTL mapping criteria is thus a promising approach for incorporating and better exploiting heterogeneous biological information into powerful and interpretable genomic prediction models

Mots clés: genomyque

Domaines: Génétique animale

Fanny Mollandin, Andrea Rau, Pascal Croiseau. Evaluating the Interpretability of SNP Effect Size Classes in Bayesian Genomic Prediction Models. 49th European Mathematical Genetics Meeting (EMGM), Apr 2021, Paris, France. pp.69-100, ⟨10.1159/000516194⟩. ⟨hal-03744745⟩