An increasingly large number of genomes is being sequenced and has to be
processed. Hence, the task of automatic genome annotation has never been
more important. We have recently developed a novel machine learning
based system for accurate computational gene finding, called mGene. The
superiority of mGene's ab initio predictions has been highlighted
by the nGASP genome annotation challenge for
nematode genomes. Despite the considerable progress initiated by
discriminative gene finders like mGene, the problem of gene finding is
still not satisfyingly solved. This is illustrated by the fact that
there is still no higher eukaryote for which all protein coding genes
are known.

Recent experimental transcriptome analysis techniques like tiling arrays
and Next Generation RNA-Sequencing (RNA-Seq) allow for high-throughput
detection of transcripts. However, the detection of precise gene
structures from these types of data alone is still very difficult. In
the case of tiling array data one has to deal with the low resolution
and the relatively high noise level. RNA-Seq provides nucleotide
resolution measurements, which are, however, unevenly distributed
over the genome.

The main result of this thesis is a gene finding system that takes full
advantage of elaborate sequence-based features as well as
high-throughput transcriptome measurements. For this we needed to
develop algorithms that are able to deal with a variety of heterogeneous
features. We designed a generic interface extending mGene that allows the
integration of tiling array as well as RNA-Seq data. It allows one to
learn to combine the different sources of information in a globally
optimal way using state-of-the-art machine learning techniques based on
Hidden Markov Support Vector Machines (HM-SVMs). Thereby, we gain the
flexibility to utilize not only transcriptome measurements, but a large
variety of data ranging from conservation scores to epigenetic features
like methylation and nuclease accessibility patterns.

We performed several experiments showing the versatility of our
approach. We particularly show that utilization of RNA-Seq as well as
tiling array data within the framework of mGene leads to substantial
improvements in gene finding accuracy.