Provenance, propagation and quality of biological annotation

Abstract

Biological databases have become an integral part of the life sciences, being used to store, organise and share ever-increasing quantities and types of data. Biological databases are typically centred around raw data, with individual entries being assigned to a single piece of biological data, such as a DNA sequence. Although essential, a reader can obtain little information from the raw data alone. Therefore, many databases aim to supplement their entries with annotation, allowing the current knowledge about the underlying data to be conveyed to a reader. Although annotations come in many di erent forms, most databases provide some form of free text annotation. Given that annotations can form the foundations of future work, it is important that a user is able to evaluate the quality and correctness of an annotation. However, this is rarely straightforward. The amount of annotation, and the way in which it is curated, varies between databases. For example, the production of an annotation in some databases is entirely automated, without any manual intervention. Further, sections of annotations may be reused, being propagated between entries and, potentially, external databases. This provenance and curation information is not always apparent to a user. The work described within this thesis explores issues relating to biological annotation quality. While the most valuable annotation is often contained within free text, its lack of structure makes it hard to assess. Initially, this work describes a generic approach that allows textual annotations to be quantitatively measured. This approach is based upon the application of Zipf's Law to words within textual annotation, resulting in a single value, . The relationship between the value and Zipf's principle of least e ort provides an indication as to the annotations quality, whilst also allowing annotations to be quantitatively compared. Secondly, the thesis focuses on determining annotation provenance and tracking any subsequent propagation. This is achieved through the development of a visualisation - i - framework, which exploits the reuse of sentences within annotations. Utilising this framework a number of propagation patterns were identi ed, which on analysis appear to indicate low quality and erroneous annotation. Together, these approaches increase our understanding in the textual characteristics of biological annotation, and suggests that this understanding can be used to increase the overall quality of these resources.