Biomedical literature mining.

Abstract

Thousands of biomedical articles are published every year containing many newly discovered biological interactions and functions. Manually reading and classifying this information is a difficult and laborious task. Literature mining contains mechanisms and tools to automate the process of extracting biological relationships, storing them in biological databases and finally analyse and present them in a biological meaningful way. In the first stage of literature mining, articles are parsed and get segmented, sentences separated, tokenized and finally annotated by part of speech tags (POS). POS tagging is the most challenging part because the training corpus is relatively small compared to the large number of biological names therefore limiting the lexicon. There are a number of solutions to address this problem including extending the lexicon manually or using character features of the word. There is no empirical comparison between different solutions. So we developed a complete list of tools including article parser, segmentation, sentence detector, sentence tokeniser, POS tagger and finally noun phrase detector using JAVA and PostgreSQL technologies. We tailored these tools for biomedical texts, and empirically compared them with other tools and we demonstrated increased efficiency of our tools compared to others. Once biological relationships are extracted they are ready to be stored in databases to be used and shared by others. There a wide range of databases that store annotation data related to genes, proteins and other biological entities. Among them Gene Ontology annotation database is the key database that connects all the other biological entities through a standard vocabulary together. In fact a Gene Ontology (GO) is a controlled vocabulary to annotate proteins based on their molecular function, biological process and cellular components. There are a number of public databases that provide data regarding GO and GO-protein relationships. We collected all relevant data from several public databases and built our specialized updatable GO database on the PostgreSQL platform. GO classification in a particular sample of genes (up/down regulated) or whole genome of a species can reveal the biological mechanisms related to its activity. Moreover, comparing the GO classification of a species under different biological conditions can elucidate its biological pathways, which can result in the discovery of novel genes to be used in therapies. We developed a web server using the PHP MVC framework connected to our specialized GO database. In this web server we developed novel visual and statistical methods to perform GO comparisons among multiple samples and genomes. We also included transcriptome based gene expression levels in GO analysis, resulting in novel meaningful biological reports. This also made comparison of whole genome gene expression across multiple biological conditions possible. Furthermore, we devised a method to dynamically construct and visualize GO regulatory networks for any gene set sample. Such a network can reveal regulatory relationships between genes helping to explain the correlated expression of genes. The topology of such a network classifies genes based on their connections, and can be used as a new method to detect important genes based on their function as well as their connectivity in the network. We demonstrated the efficiency of our developed methods in our web server by several case studies using previously published transcriptome data.