The major breakthrough at the turn of the millennium was the completion of genome sequences for many species, including humans, worms and rice. Since then, it has been realized that making use of these resources requires as much knowledge about sequences shared between members of the same species, that is, the information about sequences that distinguish individuals with different traits.

Recently, the International Rice Functional Genomics Consortium began a project to investigate sequence variation within and between rice strains on a genome-wide scale. The central goals of this project is the resequencing of the non-repetitive portion of the genomes of 20 diverse strains using high-density oligonucleotide microarrays. The resulting knowledge about sequence variation will be the basis for further evolutionary and functional studies of rice. A long-term aim is the identification of sequence variants that confer advantageous traits and that are thus targets for crop improvement through breeding.

This Diplom thesis contributed to the rice resequencing project by analysing the rich set of hybridisation data from the resequencing arrays in order to reveal hundreds of thousands of single nucleotide polymorphisms (SNPs). The application of the raw data for further computational analyses required several preprocessing steps. In a first step, we applied quantile normalisation to enable the comparison of hybridisation intensities across experiments. In a second step, we addressed the problem of cross-hybridisation by identifying repetitive 25-mers in the reference rice genome. The results of the latter analysis confirm previous reports that single nucleotide substitutions, as well as one base pair insertions/deletions (indels), have a major influence on the performance of this method.

An important part of this thesis is the compilation of a set of reference polymorphisms, which were identified by dideoxy sequencing, and to assess the quality of these reference data, which are required for subsequent SNP discovery and its evaluation. More than 3,500 fragments of average length of 557 bp had been dideoxy-sequenced from up to three strains, resulting in set of 1,755,395 positions across all strains. Polymorphic sites were be detected at 9,499 positions with 6,510 non-redundant SNP positions.

In the last step, a technique based on Support Vector Machines was applied to identify SNPs from the hybridisation data. In comparison to a model based (MB) SNP calling approach implemented by Perlegen Sciences, the machine learning (ML) method was superior by recovering 20.9% of all known SNPs at a false discovery rate (FDR) of 8.3%, compared to 13.9% and 7.8%, respectively, for the MB approach. Both methods together predicted 1,930,400 SNPs at 421,965 positions. These predictions can now be used for subsequent analyses; for example, to identify potential functional effects of the SNPs, to discover regions of the genome shared between otherwise distant strains, reflecting certain historical breeding events, or to describe genome-wide haplotype structure (i.e. construct a rice HapMap).