Mining Inter-Relationships in Online Scientific Articles and its Visualization: Natural Language Processing for Systems Biology Modeling

Abstract

With the rapid growth in the numbers of scientific publications in domains such as neuroscience and medicine, visually interlinking documents in online databases such as PubMed with the purpose of indicating the context of a query results can improve the multi-disciplinary relevance of the search results. Translational medicine and systems biology rely on studies relating basic sciences to applications, often going through multiple disciplinary domains. This paper focuses on the design and development of a new scientific document visualization platform, which allows inferring translational aspects in biosciences within published articles using machine learning and natural language processing (NLP) methods. From online databases, this software platform effectively extracted relationship connections between multiple sub-domains within neuroscience derived from abstracts related to user query. In our current implementation, the document visualization platform employs two clustering algorithms namely Suffix Tree Clustering (STC) and LINGO. Clustering quality was improved by mapping top-ranked cluster labels derived from an UMLS-Metathesaurus using a scoring function. To avoid non-clustered documents, an iterative scheme, called auto-clustering was developed and this allowed mapping previously uncategorized documents during the initial grouping process to relevant clusters. The efficacy of this document clustering and visualization platform was evaluated by expert-based validation of clustering results obtained with unique search terms.  Compared to normal clustering, auto-clustering demonstrated better efficacy by generating larger numbers of unique and relevant cluster labels. Using this implementation, a Parkinson’s disease systems theory model was developed and studies based on user queries related to neuroscience and oncology have been showcased as applications.