Large-scale automated protein function prediction
Date
2016
Authors
Kahanda, Indika, author
Ben-Hur, Asa, advisor
Anderson, Chuck, committee member
Draper, Bruce, committee member
Zhou, Wen, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
Proteins are the workhorses of life, and identifying their functions is a very important biological problem. The function of a protein can be loosely defined as everything it performs or happens to it. The Gene Ontology (GO) is a structured vocabulary which captures protein function in a hierarchical manner and contains thousands of terms. Through various wet-lab experiments over the years scientists have been able to annotate a large number of proteins with GO categories which reflect their functionality. However, experimentally determining protein functions is a highly resource-intensive task, and a large fraction of proteins remain un-annotated. Recently a plethora automated methods have emerged and their reasonable success in computationally determining the functions of proteins using a variety of data sources – by sequence/structure similarity or using various biological network data, has led to establishing automated function prediction (AFP) as an important problem in bioinformatics. In a typical machine learning problem, cross-validation is the protocol of choice for evaluating the accuracy of a classifier. But, due to the process of accumulation of annotations over time, we identify the AFP as a combination of two sub-tasks: making predictions on annotated proteins and making predictions on previously unannotated proteins. In our first project, we analyze the performance of several protein function prediction methods in these two scenarios. Our results show that GOstruct, an AFP method that our lab has previously developed, and two other popular methods: binary SVMs and guilt by association, find it hard to achieve the same level of accuracy on these two tasks compared to the performance evaluated through cross-validation, and that predicting novel annotations for previously annotated proteins is a harder problem than predicting annotations for uncharacterized proteins. We develop GOstruct 2.0 by proposing improvements which allows the model to make use of information of a protein's current annotations to better handle the task of predicting novel annotations for previously annotated proteins. Experimental results on yeast and human data show that GOstruct 2.0 outperforms the original GOstruct, demonstrating the effectiveness of the proposed improvements. Although the biomedical literature is a very informative resource for identifying protein function, most AFP methods do not take advantage of the large amount of information contained in it. In our second project, we conduct the first ever comprehensive evaluation on the effectiveness of literature data for AFP. Specifically, we extract co-mentions of protein-GO term pairs and bag-of-words features from the literature and explore their effectiveness in predicting protein function. Our results show that literature features are very informative of protein function but with further room for improvement. In order to improve the quality of automatically extracted co-mentions, we formulate the classification of co-mentions as a supervised learning problem and propose a novel method based on graph kernels. Experimental results indicate the feasibility of using this co-mention classifier as a complementary method that aids the bio-curators who are responsible for maintaining databases such as Gene Ontology. This is the first study of the problem of protein-function relation extraction from biomedical text. The recently developed human phenotype ontology (HPO), which is very similar to GO, is a standardized vocabulary for describing the phenotype abnormalities associated with human diseases. At present, only a small fraction of human protein coding genes have HPO annotations. But, researchers believe that a large portion of currently unannotated genes are related to disease phenotypes. Therefore, it is important to predict gene-HPO term associations using accurate computational methods. In our third project, we introduce PHENOstruct, a computational method that directly predicts the set of HPO terms for a given gene. We compare PHENOstruct with several baseline methods and show that it outperforms them in every respect. Furthermore, we highlight a collection of informative data sources suitable for the problem of predicting gene-HPO associations, including large scale literature mining data.