Computational Biology and Machine Learning

Gregor Mendel often called the "father of modern genetics" paved the way with his studies of pea plants for a genetically understanding of heredity and phenotypic variations of plants and animals. Roughly speaking, he demonstrated that genes are responsible for the phenotypic appearance of organisms or in its simplest form that one genes determines a trait, e.g., the color of a seed, completely. Such genes are called Mendelian genes. Despite the success to explain with Mendelian genes many observable phenotypes there are severe limitations especially regarding the understanding of diseases, which can be seen as phenotype of organisms. It has been realized that many diseases, such as diabetes, heart disease or cancer, depend on altered interactions between multiple genes, rather than changes in a single gene. This means, that groups of interacting genes, frequently called pathways, seem to be more appropriate for a systematic analysis and, hence, understanding of such complex diseases.

The problem we are facing with complex diseases is visualized in the figure above. Due to the lack of simple biomarkers we can not even reliably detect the presence and the type of a complex disease. This means, a direct comparison on the patients level based on simple biomarkers is not possible. For this reason, we are working around this obstacle by deriving a representation from experimental data that is complex enough to capture relevant molecular interactions or modifications thereof that allows us to detect the presence, absence or the type of a comple disease on the level of this mathematical representation. We use high-throughput data from microarray experiments from which we are deriving pseudo pathways of biological processes that can be compared graph theoretically. Due to the fact, that the infered pseudo pathways are erroneous and large such a comparison is not straight forward but needs to be done statistically.

In addition to this approach we are currently developing another method that aims to measure the interaction structure on the expression level directly.

We developed a graph theoretical method that allows to make predictions about the functional annotation of genes. Our approach is based on a given causal network structure of a gene network and forms an extension of the well-known measure betweenness centrality originally developed to study social networks. We call our measure joint betweenness because it evaluates the joint occurance of nodes on shortest communication paths found in the network. For the transcriptional regulatory network of Saccharomyces cerevisiae we demonstrated that our method works statistically by using a test set of annotated genes. Further, we made predictions for some unannotated genes that appear to be plausible because all predicted genes are involved in metabolic pathways and, hence, form a coherent subgraph of the overall transcriptional regulatory network.

I am interested in the organizational principles of the protein structure as well as in the dynamical interplay between genes regulating and controlling the machinery within a cell, which is the smallest entity of life, to ensure and maintain an appropriate functioning of an organism. In one of my current projects, I try to determine the structural domains of a protein. If the structural organization of a protein is viewed hierarchically, a domain of a protein is above the amino acid and secondary structure element level, but below the tertiary structure of the protein. See the example shown in the figures for a two domain protein, PDB ID 1OE1, according to the assignment by SCOP. The identification of structural domains of a protein is important, because a meaningful domain identification enables a structural comparison and classification of protein domains. This would complement methods for sequence comparison and may reveal insightes into, e.g., questions concerning evolution, sequence analysis alone may not be able to provide. A reason therefore is, that during evolution the structure of a protein is more conserved than its amino acid sequence.