Protein solubility classification in biomedical concept space : doktor of philosophy thesis
Klasifikacija topnosti proteinov v prostoru biomedicinskih konceptov tekstovne analize : doktorska disertacija

english

doktorska dela

2011

Univerza v Mariboru, Fakulteta za elektrotehniko računalništvo in informatiko

Proteins are an essential part of every organism and each protein has its own function, which depends on the proteinćs structure. The latter is an importantresearch topic and researchers often isolate proteins from complex mixtures to study their structures. The isolation process is in many ways influenced by the proteinćs solubility since insoluble proteins are usually harder to isolate than soluble ones. In addition, low protein solubility has been linked to different diseases. For these reasons, researchers often wish to indentify which proteins are more likely to be soluble. As a result, several protein solubility classification algorithms have been proposed. Roughly speaking, these algorithms take a set of soluble and insoluble proteins as an input, learn their differences and product a classifier that can be used to predict solubility for new proteins. In this thesis we propose a new method for protein solubility classification, which uses text mining techniques to define protein attributes. This new method extracts biomedical knowledge from scientific literature and presents this knowledge in the form of so called biomedical concept attributes. These attributes present a novel approach of describing proteins in the classification process, since todayćs state-of-the-art classification methods use mostly attributes derived from theproteinćs sequence. To evaluate the new method, this thesis describes the classification scheme for an empirical study which measures the impact of the new attributes on the protein solubility classification. In the study, the twenty most common sequence derived attribute datasets are analysed, to which we gradually add five types of biomedical concept attributes. We measure the performance of the classifiers obtained by these attribute datasets. As a result, this thesis introduces several original scientific contributions. First of all, an analysis of protein databases that contain information about protein solubility is performed. Secondly, the method for extracting biomedical concept attributes is presented. Next, an original comparison of methods that use biomedical concept attributes with those that use only sequence-derived attributes is performed. The thesis demonstrates that the newattributes increase the performance of some classifiers. Finally, it identifies types of words and word associations from the medical literature that are associated with protein solubility.

07.08.2023