Information Extraction from Text for Improving Research on Small Molecules and Histone Modifications

The cumulative number of publications, in particular in the life sciences, requires efficient methods for the automated extraction of information and semantic information retrieval. The recognition and identification of information-carrying units in text – concept denominations and named entities – relevant to a certain domain is a fundamental step. The focus of this thesis lies on the recognition of chemical entities and the new biological named entity type histone modifications, which are both important in the field of drug discovery. As the emergence of new research fields as well as the discovery and generation of novel entities goes along with the coinage of new terms, the perpetual adaptation of respective named entity recognition approaches to new domains is an important step for information extraction. Two methodologies have been investigated in this concern: the state-of-the-art machine learning method, Conditional Random Fields (CRF), and an approximate string search method based on dictionaries. Recognition methods that rely on dictionaries are strongly dependent on the availability of entity terminology collections as well as on its quality. In the case of chemical entities the terminology is distributed over more than 7 publicly available data sources. The join of entries and accompanied terminology from selected resources enables the generation of a new dictionary comprising chemical named entities. Combined with the automatic processing of respective terminology – the dictionary curation – the recognition performance reached an F1 measure of 0.54. That is an improvement by 29 % in comparison to the raw dictionary. The highest recall was achieved for the class of TRIVIAL-names with 0.79. The recognition and identification of chemical named entities provides a prerequisite for the extraction of related pharmacological relevant information from literature data. Therefore, lexico-syntactic patterns were defined that support the automated extraction of hypernymic phrases comprising pharmacological function terminology related to chemical compounds. It was shown that 29-50 % of the automatically extracted terms can be proposed for novel functional annotation of chemical entities provided by the reference database DrugBank. Furthermore, they are a basis for building up concept hierarchies and ontologies or for extending existing ones. Successively, the pharmacological function and biological activity concepts obtained from text were included into a novel descriptor for chemical compounds. Its successful application for the prediction of pharmacological function of molecules and the extension of chemical classification schemes, such as the the Anatomical Therapeutic Chemical (ATC), is demonstrated. In contrast to chemical entities, no comprehensive terminology resource has been available for histone modifications. Thus, histone modification concept terminology was primary recognized in text via CRFs with a F1 measure of 0.86. Subsequent, linguistic variants of extracted histone modification terms were mapped to standard representations that were organized into a newly assembled histone modification hierarchy. The mapping was accomplished by a novel developed term mapping approach described in the thesis. The combination of term recognition and term variant resolution builds up a new procedure for the assembly of novel terminology collections. It supports the generation of a term list that is applicable in dictionary-based methods. For the recognition of histone modification in text it could be shown that the named entity recognition method based on dictionaries is superior to the used machine learning approach. In conclusion, the present thesis provides techniques which enable an enhanced utilization of textual data, hence, supporting research in epigenomics and drug discovery