Date of Award:


Document Type:


Degree Name:

Doctor of Philosophy (PhD)


Computer Science


Dr. Xiaojun Qi


In recent years, machine learning has emerged as an important discipline. However, despite the popularity of machine learning techniques, data in the form of discrete structures are not fully exploited. For example, when data appear as graphs, the common choice is the transformation of such structures into feature vectors. This procedure, though convenient, does not always effectively capture topological relationships inherent to the data; therefore, the power of the learning process may be insufficient. In this context, the use of kernel functions for graphs arises as an attractive way to deal with such structured objects.

On the other hand, several entities in computational biology applications, such as gene products or proteins, may be naturally represented by graphs. Hence, the demanding need for algorithms that can deal with structured data poses the question of whether the use of kernels for graphs can outperform existing methods to solve specific computational biology problems. In this dissertation, we address the challenges involved in solving two specific problems in computational biology, in which the data are represented by graphs.

First, we propose a novel approach for protein function prediction by modeling proteins as graphs. For each of the vertices in a protein graph, we propose the calculation of evolutionary profiles, which are derived from multiple sequence alignments from the amino acid residues within each vertex. We then use a shortest path graph kernel in conjunction with a support vector machine to predict protein function. We evaluate our approach under two instances of protein function prediction, namely, the discrimination of proteins as enzymes, and the recognition of DNA binding proteins. In both cases, our proposed approach achieves better prediction performance than existing methods.

Second, we propose two novel semantic similarity measures for proteins based on the gene ontology. The first measure directly works on the gene ontology by combining the pairwise semantic similarity scores between sets of annotating terms for a pair of input proteins. The second measure estimates protein semantic similarity using a shortest path graph kernel to take advantage of the rich semantic knowledge contained within ontologies. Our comparison with other methods shows that our proposed semantic similarity measures are highly competitive and the latter one outperforms state-of-the-art methods. Furthermore, our two methods are intrinsic to the gene ontology, in the sense that they do not rely on external sources to calculate similarities.


This work made publicly available electronically on April 12, 2012.