Homologousproteins do not necessarily exhibit identical biochemicalfunction. Despitethis fact, local or global sequence similarity is widely used as an indication of functional identity. Ofthe 1327Enzyme Commissiondefined functional classes with morethan one annotated example in the sequence databases, similarity scores alone are inadequate in 251 (19%) of the cases. Wetest the hypothesis that conserved domains, as defined in the ProDomdatabase, can be used to discriminate betweenalternative functions for homologousproteins in these cases. Using machine learning methods, wewere able to induce correct discriminators for morethan half of these 251 challenging functional classes. Theseresults showthat the combination of modular representations of proteins with sequencesimilarity improvesthe ability to infer function fromsequenceover similarity scores alone.