Abstract :
[en] The recent developments in molecular biology have made available thousands of genetic markers, allowing livestock genotyping at a reasonable cost and the subsequent development of genomic prediction. The single-step procedure, a unified approach of genomic prediction, requires inversion of two matrices gathering additive relationships between genotyped animals: the genomic relationship matrix (G) and a part of the additive relationship matrix (A22). The inverse of A22 may also be interesting for other applications. Matrix inverse can be constructed successively by, first, computing, for each animal, the vector containing contributions of other animals to its relationship and, secondly, adding the product of each vector of contributions by itself to a zeroed matrix. The objectives of this thesis were (1) to propose algorithms to compute or to approximate the vector of contributions and (2) to test the numerical efficiency of these algorithms (computing speed, memory use and, if needed, approximation accuracy). Computing contributions covered two points: (1) finding or approximating which contributions are different from zero, and (2) computing the value of contributions considered as non-zero. In the first approach, we considered that animals closely related have non-zero contributions and approximated their values by linear regression. This approach was extended in a recursive way. In the second approach, we empirically determined the set of non-zero contributions by a heuristic algorithm of pedigree exploration (only for the case of A22). Values were then computed either by linear regression, or using the already computed inverse. We also tested an approximation strategy: limiting the number of extracted generations of non-genotyped ancestors to reduce pedigree complexity. In a third approach, we followed the same heuristic algorithm as before but restricted the pedigree exploration to find out which animals have a non-zero contribution. Their values were approximated by linear regression. The presentation of the different approaches is followed by a general discussion in which the approaches are compared. It was found that the best compromise between speed, memory and approximation accuracy was achieved by the last approach for the case of A22. Use of this last approach simplified computations and therefore made predictions more feasible. However, for the case of G, no sufficient approximations could be reach in a reasonable time. Perspectives of other uses of algorithms developed and of future researches were drawn, as well as practical perspectives for animal breeding.
