Three-step Bayesian factor analysis applied to QTL detection in crosses between outbred pig populations

Abstract

Marker assisted selection (MAS) can be used to improve the efficiency of genetic selection of traits for which phenotypic measurements are expensive or cannot be obtained on selection candidates, such as carcass traits. Marker information required for MAS may be acquired through the identification of QTLs. Generally, univariate models are used for QTL detection, although multiple-trait models (MTM) may enhance QTL detection and breeding value estimation. In MTM, however, the number of parameters can be large and, if traits are highly correlated, such as carcass traits, estimates of (co)variance matrices may be close to singular. Because of this, dimension reduction techniques such as Factor Analysis (FA) may be useful. The aim of our project is to evaluate the use of FA for structuring (co)variance matrices in the context of Bayesian models for QTL detection in crosses between outbred populations. In our method, QTL effects are postulated at the level of common factors (CF) rather than the original traits, using a three-step approach. In a first step, a MTM is fitted to arrive at estimates of systematic effects and prediction of breeding values (procedure A) and only systematic effect (procedure B). These estimates/predictions are then used to generate an adjusted phenotype that is further analyzed with a Bayesian FA model. This step yields estimates of factor scores for each animal and CF. In the last step, the scores relative to each CF are analyzed independently using probabilities for the line of origin combination. To illustrate the methodology, data on 416 F2 pigs (Brazilian Piau X commercial) with ten traits (5 fat-related, 2 loin measurements, and 3 carcass classification systems) were analyzed. For each of the three resulting CFs, an independent QTL scan was performed on chromosome 7 considering three models: I) null (i.e., absence of QTL); II) additive effect QTL, and III) additive and dominance effect QTL. The posterior probability (PP) of each model was calculated from Bayes factor for each considered procedures (A and B). A Three-step Bayesian factor analysis allowed us to calculate the probability of QTLs that simultaneously affect a group of carcass traits for each position of SSC 7. The removal of systematic effects in the first step of the evaluation (procedure B) allowed that the factor analysis, which was performed in the second step, identify three distinct factors that explained 85% of the total traits variation. For the common factor that represented fat-related traits (bacon depth, midline lower backfat thickness, higher backfat thickness on the shoulder; midline backfat thickness after the last rib; midline backfat thickness on the last lumbar vertebrae) the third step of the analysis showed that the highest probability of an additive QTL effect at the 65cM position was 86%.