The applications of graphs in cheminformatics are numerous. From modeling molecules themselves to reaction networks, to ring system indicies, and even scientific paper authorship relationships, graphs are the way most problems get framed.
Wouldn't it be great if your natural mental model of a problem didn't have to change just because you needed persistent storage of the underlying data?
Neo4j, a high-performance, transactional, embedded graph database written in Java, offers what appears to be the best solution to this gnarly problem I've ever seen.
Check out the presentation by Emil Eifrem and see for yourself.