Concept profiles for genes are built from the relative frequency of words and terms in PubMed abstracts. In turn, overlap of the profiles are used to identify gene-gene relationships. In practice, the procedure provides intuitively reasonably results, but there is no way of rigorously benchmarking such knowledge generated networks. The method occasionally errs on the side of over-inclusiveness. For example, it is not able to distinguish between statements such as 'protein A is associated with disease B' versus 'protein A is not associated with disease B'. As illustrated in the Results, it is also possible for a disease and gene to be linked by irrelevant factors, such as symptoms common to more than one syndrome. Similarly, gene-gene relationships may sometimes be based on non-pathway related factors. For example the 13 members of the human kallikrein family are tightly coupled, because of many articles that discuss them as a group. In fact, most of the family members operate in quite different pathways. In future, more sophisticated natural language processing technology may be applied to reduce these effects. At present, a concept
