Construction and computation methods for biological networks

Abstract

Biological systems are complex in that they comprise large number of interacting entities, and their dynamics follow mechanic regulations for movement and biological function organization. Established computational modeling deals with studying and manipulating biologically relevant systems as a powerful approach. Inner structure and behavior of complex biological systems can be analyzed and understood by computable biological networks. In this thesis, models and computation methods are proposed for biological networks.

The study of Genetic Regulatory Networks (GRNs) is an important research topic in genomic research. Several promising techniques have been proposed for capturing the behavior of gene regulations in biological systems. One of the promising models for GRNs, Boolean Network (BN) has gained a lot of attention. However, little light has been shed on the analysis of internal connection between the dynamics of biological molecules and network systems. Inference and completion problems of a BN from a given set of singleton attractors are considered to be important in understanding the relationship between dynamics of biological molecules and network systems. Discrete dynamic systems model has been recently proposed to model time-course microarray measurements of genes, but delay effect may be modeled as a realistic factor in studying GRNs. A delay discrete dynamic systems model is developed to model GRNs.

Inference and analysis of networks is one of the grand challenges in modern statistical biology. Machine learning method, in particular, Support Vector Machine (SVM), has been successfully applied in predictions of internal connections embedded in networks. Kernels in conjunction with SVM demonstrate strong ability in performing various tasks such as biomedical diagnosis, function prediction and motif extractions. In biomedical diagnosis, data sets are always high dimensional which provide a challenging research problem in machine learning area. Novel kernels using distance-metric that are not common in machine learning framework are proposed for possible tumor differentiation discrimination problem.

Protein function prediction problem is a hot topic in bioinformatics. The K-spectrum Kernel is among the top popular models in description of protein sequences. Taking into consideration of positive-semi-definiteness in kernel construction, Eigen-matrix translation technique is introduced in novel kernel formulation to give better prediction result. In a further step, power of Eigen-matrix translation technique in feature selection is demonstrated through mathematical formulation. Due to structure complexity of carbohydrates, the study of carbohydrate sugar chains has lagged behind compared to that of DNA and proteins. A weighted q-gram kernel is constructed in classifying glycan structures with limitations in feature extractions. A biochemically-weighted tree kernel is then proposed to enhance the ability in both classification as well as motif extractions.

Finally the problem of metabolite biomarker discovery is researched. Human diseases, in particular metabolic diseases, can be directly caused by the lack of essential metabolites. Identification of metabolite biomarkers has significant importance in the study of biochemical reaction and signaling networks. A promising computational approach is proposed to identify metabolic biomarkers through integrating biomedical data and disease-specific gene expression data.