The challenge and limitations of combining data: a case study examining the relationship between intramuscular fat content and flavour intensity based on the BIF-BEEF database
Résumé
The BIF-BEEF (Beef Integrated and Functional Biology) data warehouse for muscle biology to predict beef quality gathers data related to bovines, their carcasses and their beef. These data come mainly from three sources: the INRA database named FiLiCol, the European GEMQUAL program and the French QUALVIGENE program databases plus other minor sources. At the beginning of 2011, the BIF-BEEF data warehouse contained 331 745 measurements for 621 variables. Measurements were obtained on eight muscles and/or from 5197 animals (mainly young bulls) belonging to 20 different breeds (mainly Charolais, Limousin, Blonde d'Aquitaine, the three major French beef breeds) from experiments carried out over a 10-year period. A web interface was developed to extract data and to analyse them using basic statistical tools (correlation, variance analysis, etc) with R software. Clearly, since the various experiments were not designed initially to ultimately link together, it appeared very difficult to integrate some data which differ a lot by units, scales or laboratory methods. Ontology will help to address these issues. However, the usefulness of the BIF-BEEF data warehouse is described by studying the relationship in M. longissimus thoracis between intramuscular fat content (IMF) and flavour assessed by sensory panels. When data from different sources or different sensory panels were used, they were corrected for these fixed factors in the regression model. They were also corrected for known sources of variation (sex, breed and age of the animals). On average, the relationship between IMF and flavour was low (partial correlation coefficient r = 0.11) but significant. This relationship was no more significant for breeds with low IMF levels (such as Blonde d'Aquitaine) or for animals with the highest IMF such as steers or females.