One of the most important use cases of ontologies is the calculation of similarity scores between a query and items annotated with classes of an ontology. The hierarchical structure of an ontology does not necessarily reflect all relevant aspects of the domain it is modelling, and this can reduce the performance of ontology-based search algorithms. For instance, the classes of phenotype ontologies may be arranged according to anatomical criteria, but individual phenotypic features may affect anatomic entities in opposite ways. Thus, "opposite" classes may be located in close proximity in an ontology; for example enlarged liver and small liver are grouped under abnormal liver size. Using standard similarity measures, these would be scored as being similar, despite in fact being opposites. In this paper, we use information about opposite ontology classes to extend two large phenotype ontologies, the human and the mammalian phenotype ontology. We also show that this information can be used to improve rankings based on similarity measures that incorporate this information. In particular, cosine similarity based measures show large improvements. We hypothesize this is due to the natural embedding of opposite phenotypes in vector space. We support the idea that the expressivity of semantic web technologies should be explored more extensively in biomedical ontologies and that similarity measures should be extended to incorporate more than the pure graph structure defined by the subclass or part-of relationships of the underlying ontologies.