The uncovering of novel relationships between biomedical entities (e.g., diseases, chemicals, genes, proteins, metabolites, pathways) is of great importance in improving our understanding of biology and in developing new and useful therapeutics. A number of efforts have previously sought to identify novel relationships from automated text mining of research literature. The task is challenging, however, as indicated by the performances of 18 different teams who participated in a competition to extract chemical-disease relationships from text at the 2015 BioCreative Workshop (F-score ranges of 32.01% to 57.03%). Inspired by the innovative and successful approach by Spangler et. al to predict kinase interactions with the protein TP53 from mining scientific literature, we have developed a novel approach for mining relationships from both scientific abstracts and full text research articles. Our approach extends upon our previous development of Chemotext and builds upon development of word embedding models combined with graph-based similarity algorithms to generate a predicted interaction score between pairs of biomedical entities. We have initially applied our approach to the task of identifying novel kinase-TP53 interactions. Using time-split validation, our top set of predicted novel kinase-TP53 interactions include two of those that were validated later. Of note, our top predictions included the top prediction from the Spangler et al paper that was later validated in the laboratory as part of that effort. However, in addition, our approach gave low scores to kinases that were predicted to interact with TP53 in the Spangler paper but that were not validated in subsequent laboratory tests. We will discuss ongoing studies to improve and expand our approach, both for protein-protein interactions and for the more general case of chemical-target-gene interactions.

Mining protein interactions from biomedical literature using semantic similarity

253rd National Meeting (2017)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.