Variance components estimation of complex traits including microbiota information

The influence of the microbiome on relevant complex traits for dairy cattle, such as feed efficiency or methane emissions has been well established. Further, recent studies have released evidences on the control of the genetic background of the animal over the microbiota composition. However, until now most analyses have focused on single microorganism approaches instead of the joint microbiome as a whole, including underlying relationships. The joint analysis of the genetic background of the host and its microbiota requires accounting for the distance (or dissimilarity) between communities of microorganisms in different hosts. Therefore, it is necessary to incorporate the whole microbiome into the statistical models to assess its association with complex traits. Microbiome relationship matrix (MRM) allow considering the microbiota as a whole. Several methods have been proposed to ordinate these matrices; those differ on the metric used to account for the distance (or dissimilarities) between microbial communities (e.g. Euclidean, Bray-Curtis, χ2). These distances account for alpha and beta diversity in different ways. Consensus on what method is the most appropriate hasn’t been reached yet, and might depend on data singularities and the purpose of the study. The aim of this study was to compare several microbiota relationship matrices, within a variance component estimation framework. Five ordination methods to build the MRM were tested: metric multidimensional scaling (MDS), detrended correspondence analysis (DCA), non-metric multidimensional scaling (NMDS), redundancy analysis (RDA) and constrained correspondence analysis (CCA). The log transformed and standardized relative abundances matrix described in Ross et al. (2013) was used as a benchmark matrix. Simulated (n equal to 1000) data were used to estimate variance components including phenotypes, genotypes and rumen microbiota information. Data were analysed considering two possible models. First, the genomic effect and the microbiota effect were included independently. Second, an interaction effect between the genomic and microbiota effects was added. All models were implemented within a Bayesian framework using the BGLR package in R. A total of 100 replicates were generated. Real data were analysed using the same models. Similar or slightly better estimation of simulated h2 (0.30) and m2 (0.50) in the independent effects models resulted from ordination methods of MDS (0.307 and 0.493), RDA (0.307 and 0.501) and CCA (0.305 and 0.500) compared to the benchmark MRM (0.304 and 0.480), while poor performance of the DCA (0.249 and 0.349) and NMDS (0.217 and 0.266) methods were obtained at estimating those parameters. The correlation coefficients between genomic estimated breeding values (GEBV) and true breeding values (TBV), from higher to lower, were: the obtained with the benchmark matrix (rho equal to 0.633), CCA (rho equal to 0.631), RDA (rho equal to 0.624), DCA (rho equal to 0.598), MDS (rho equal to 0.592) and NMDS (0.557). Likewise, correlations for predicted microbiota effect in the same order were: the benchmark matrix (rho equal to 0.975), CCA (rho equal to 0.966), RDA (rho equal to 0.949), MDS (rho equal to 0.845), DCA (rho equal to 0.807) and NMDS (rho equal to 0.517). Similar results, in terms of matrices performance, were obtained for the interaction effects model. A real data set (n equal to 70) was also analysed under the same frameworks. Low heritability estimates for feed efficiency (from 0.077 to 0.083) and microbiability (from 0.073 to 0.103) were observed; however, consistent values for the microbiability were obtained with the MRM that performed better in the simulations (from 0.073 to 0.077). Besides, high correlations (rho greater than 0.85) between the genetic effect of the host and the phenotypes were obtained for all methods, as well as high correlations between the microbiota effect and the phenotypes for the RDA (rho equal to 0.91) and CCA (rho equal to 0.91) matrices. Both models were compared using the deviance information criteria (DIC), the effective number of parameters (pD), and the posterior mean of the log likelihood (PostMeanLogLik), resulting in slightly lower values for the independent effects model (DIC: 183.9 to 189.3) than the interaction effects model (DIC: 187.5 to 191.7), those results indicate that it might be a relationship linking genotype-microbiome-phenotype which could be used in prediction of complex traits. The analyses performed in this thesis suggest that canonical ordination methods of RDA and CCA to create MRM are preferred when whole microbiota information is included in the statistical models to analyse complex traits