Browsing by Author "Zosa, Elaine"

Protein function prediction aims to identify the function of a given protein using, for example, sequence data, protein-protein interaction or evolutionary relationships. The use of biomedical literature to predict protein function, however, is a relatively under-studied topic given the vast amount of readily available data. This thesis explores the use of abstracts from biomedical literature to predict protein functions using the terms specified in the Gene Ontology (GO). The Gene Ontology (GO) is a standardised method of cataloguing protein functions where the functions are organised in a directed acyclic graph (DAG). The GO is composed of three separate ontologies: cellular component (CC), molecular function (MF) and biological process (BP). Hierarchical classification is a classification method that assigns an instance to one or more classes where the classes are hierarchically related to each other, as in the GO. We build a hierarchical classifier that assigns GO terms to abstracts by training individual binary Naïve Bayes classifiers to recognise each GO term. We present three different methods of mining abstracts from PubMed. Using these methods we assembled four datasets to train our classifiers. Each classifier is tested in three different ways: (a) in the paper-centric approach, we assign GO terms to a single abstract, (b) in the protein-centric approach, we assign GO terms to a concatenation of abstracts relating to single protein; and (c) the term-centric approach is a complement of the protein-centric approach where the goal is to assign proteins to a GO term. We evaluate the performance of our method using two evaluation metrics: maximum F-measure (F-max) and minimum semantic distance (S-min). Our results show that the best dataset for training our classifier depends on the evaluation metric, the ontology and the proteins being annotated. We also find that there is a negative correlation between the F-max score of a GO term and its information content (IC) and a positive correlation between the F-max and the term's centrality in the DAG. Lastly we compare our method with GOstruct, the state-of-the-art literature-based protein annotation program. Our method outperforms GOstruct on human proteins, showing a significant improvement for the MF ontology.