Extracting information for biology

High-throughput methods like the large scale sequencing of the human genome dramatically increase our knowledge of genetics and related biological processes. As a consequence these results accelerate the pace of research and development in the field of biomedicine. The overall goal of these research efforts is to obtain new findings about diseases in order to improve human health. However, these advances are responsible for an increase in complexity and a need for understanding when applying biomedical research and data. Meanwhile there is a strong agreement within life-science related academic laboratories and industry that addressing the complexity of biological data and knowledge entails intense interdisciplinary efforts. A major requirement for interdisciplinary research within life sciences is to correlate the data that is derived from text with data from experiments in biomedical laboratories (and with patient records). The main contribution of this work is to describe how natural language processing (NLP) methods and systems can fulfill this requirement by categorising, structuring and exploiting the massive amount of textual data available and in integrating the results with data derived from biomedical experiments. The present work is thematically divided into three parts. The first part is about text mining in the life sciences and is subdivided in two subsections. Subsection I presents an introduction to effective natural language processing techniques for identifying and retrieving information from large text collections. Furthermore it presents the characteristic features of biomedical terminology, which comprise synonymic, homonymic, orthographical, paragrammatical as well as other types of variance. This illustrates that the crucial difference between everyday language and the language used within biomedical scientific literature is mainly based on the difference of the terminology used. This subsection concludes with a description of basic criteria that an information extraction system has to meet. The implementation of such an information extraction system is described in the second subsection. This section documents a pilot study that was carried out in close collaboration with both the SDBV (Scientific DataBases and Visualisation) group of EML Research gGmbH and Peer Bork's group at the EMBL (European Molecular Biology Laboratory) both located in Heidelberg. The system implemented is used for the extraction of information on gene expression relations from biomedical scientific publications. The second part III focuses on the transfer of a computational linguistic tool (TIGERSearch), which was originally developed for the querying of hierarchical structures, to querying knowledge on protein domains from a protein database. It is demonstrated that TIGERSearch offers the possibility to make implicit knowledge about protein domains explicit by transforming the database entries to TIGERSearch-XML. In addition, TIGERSearch makes this implicit knowledge graphically visible. In fact, TigerSearch was initially developed for the querying and transparent representation of syntactically annotated corpora, so-called treebanks. This part also points out the problem that mapping the wide range of natural language annotations to precisely defined concepts presupposed by the search engine requires an ontological modelling of the domain. The third part addresses the problem of ontological modelling in a more general and more comprehensive way. It consists of two chapters. The first chapter introduces basic notions of ontologies as well as an overview of guidelines to be considered when building an ontology. In addition some examples of implemented (both general and biomedical) ontologies are presented. The second chapter presents an axiomatisation of a sub-domain of molecular biology (i.e. gene expression) that comprises the domain of proteins and their domains. The thesis demonstrates a highly interdisciplinary approach for text mining in the life sciences. Methods and knowledge from the fields of natural language processing, bioinformatics and biology have been successfully combined with knowledge from cell-biology and the problem of extracting knowledge from unstructured or partially structured data. Im Rahmen der vorliegenden Arbeit habe ich Methoden der Computerlinguistik diskutiert, erarbeitet und eingesetzt, um eine maschinelle Extraktion biomedizinischer Daten aus wissenschaftlichen Publikationen und Datenbanken zu ermoglichen. Im wesentlichen habe ich dabei drei Themengebiete bearbeitet. Im ersten Teil der Arbeit stelle ich eine Pilotstudie vor, die die Methoden der Informationsextraktion anwendet, um fur eine vordefinierte Fragestellung Antworten aus grosen Mengen biomedizinischer Texte zu extrahieren. Diese Arbeit ist von unmittelbarer biologischer Relevanz. Denn nachdem diese Arbeit in enger Kollaboration zwischen der SDBV (Scientific DataBases and Visualisation) Gruppe der EML Research gGmbH und der Gruppe um Peer Bork des EMBL (Europaisches Molekularbiologie Labor) entstanden ist, wird das System zur Extraktion von Genexpressionsdaten am EMBL eingesetzt. Im zweiten Teil der Arbeit stelle ich Einsatzmoglichkeiten der TigerSearch Suchmaschine im molekularbiologischen Kontext vor. TigerSearch wurde fur die Suche auf syntaktisch annotierten Satzen entwickelt. Ich habe sie ausgewahlt, um strukturierte Information uber Proteindomanen transparent darzustellen und komplexe Anfragen bearbeiten zu konnen. Der letzte Teil der Arbeit beschaftigt sich mit der Modellierung biologischen Wissens. Dabei steht die formale Reprasentation biologisch relevanter Konzepte im Vordergrund. Als wesentliches Merkmal der vorliegenden Arbeit mochte ich den interdisziplinaren Charakter betonen, vor allem, weil interdisziplinare Forschung nicht nur in der Bioinformatik, sondern auch in anderen Forschungsfeldern zunehmend an Bedeutung gewinnt.