Up to now, the emphasis of the text-mining techniques employed by LBD tools has been on the correct identification of biomedical terms in text. Most approaches use the UMLS Metathesaurus as the source of terms. When deploying LBD in a genomics context, this may be a limiting factor because the coverage of gene and proteins in the UMLS is low. [...] Unambiguously recognising gene symbols, however, is not straightforward as many gene symbols refer to one or more genes or other meanings. Wren partly solves this in IRIDESCENT by matching the gene’s acronym to its accompanying full name. Literature-based discovery tools should benefit from recent research in identifying gene names41 and their disambiguation.

Not only are text-mining and information extraction tools needed to render the biomedical literature accessible but the results of these tools can also assist researchers in the formulation and evaluation of novel hypotheses. This requires an additional set of technological approaches that are defined here as literature-based discovery (LBD) tools. [...] The process of (semi-)automatically inferring implicit knowledge from literature databases, which results in well- motivated and testable hypotheses, is called literature-based discovery (LBD). In this paper, a short background on LBD and an overview of recent discoveries are first provided. [...] General text-mining approaches focus on the extraction of relevant entities, eg genes and proteins, and relationships between them, eg protein– protein interactions. The users of these approaches generally retrieve known, explicit co-occurrence-based knowledge that they personally were not always aware of and text mining can be viewed as an efficient way of keeping abreast with the most important facts in the literature. Building largely on mined entities and facts, LBD tools attempt to combine extracted information into serendipitous and truly novel hypotheses. Because many combinations of mined facts are possible, the main aim of LBD systems is to confine the explosively growing number of possible hypotheses to those that have the highest probability to be consistent. Most systems provide a list of hypotheses rank-ordered according to certain likelihood, for instance, the number of intermediate B concepts of a certain AC hypothesis. In contrast to text-mining and information extraction approaches, there is no straightforward evaluation possible as it is not easy to establish the correctness of generated hypotheses.