Manually created large scale ontologies are useful for organizing, searching, and repurposing content ranging from scientific papers and medical guidelines to images. However, maintenance of such ontologies is expensive. In this paper, we investigate the use of universal schemas (Riedel et al., 2013) as a mechanism for ontology maintenance. We apply this approach on top of two unique data sources: 14 million full-text scientific articles and chapters, plus a 1 million concept handcurated medical ontology. We show that using a straightforward matrix factorization algorithm one can achieve 0.7 F1 measure on a link prediction task in this environment. Link prediction results can be used to suggest new relation types and relation type synonyms coming from the literature as well as predict specific new relation instances in the ontology.