Machine Learning for Genomic Sequence Analysis - Dissertation -

Acknowledgements Above all, I would like to thank Dr. Gunnar Rätsch and Prof. Dr. Klaus-Robert Müller for their guidance and inexhaustible support, without which writing this thesis would not have been possible. All of the work in this thesis has been done at the Fraunhofer Institute FIRST in Berlin and the Friedrich Miescher Laboratory in Tübingen. I very much enjoyed the inspiring atmosphere in the IDA group headed by K.-R. Müller and in G. Rätsch's group in the FML. Much of what we have achieved was only possible in a joint effort. As the various fruitful within-group collaborations expose — we are a team. reading the draft, helpful discussions and moral support. I acknowledge the support from all members of the IDA group at TU Berlin and Fraunhofer FIRST and the members of the Machine Learning in Computational Biology group at the Friedrich Miescher Laboratory, especially for the tolerance of letting me run the large number of compute-jobs that were required to perform the experiments. Finally, I would like to thank the system administrators became utterly excited about this subject. During that seminar, I gave an introductory talk on SVMs to an all-knowing audience, but it took a while before I seriously picked up the SVM subject in 2002. For my student research project, Gunnar Rätsch suggested Hidden Markov Models (HMMs) and their application to Bioinformatics. The idea was to find genes, but as this problem is too complex to start with, I started with the sub-problem of recognising splice sites. HMMs using a certain predefined topology performed quite well. However, we managed to improve on this using string kernels (derived from HMMs) and Support Vector Machines (Sonnenburg et al., 2002, Tsuda et al., 2002b). Unfortunately, training SVMs with these string kernels was computationally very costly. Even though we had a sample of more than 100,000 data points, we could barely afford training on 10,000 instances on a, at that time, cutting-edge Compaq Alpha compute cluster. In contrast to HMMs, the resulting trained SVM classifiers were not easily accessible. After a research break of about 2 years during which I became the group's system administrator, I started work on exactly these — still valid — topics. All of this research was application driven: improving the splicing signal (and later, other signals) detection in order to construct a gene finder that is applicable on a genomic scale. To this end, I …