In the domain of bioinformatics, the role played in the biological process by proteins, that act as transmitters and receivers of information thus ruling the mechanisms that determine how organic systems function, has great importance. Recent studies produced evidence of a strict correlation between the surface characteristics of proteins and the way they interact. In this paper we propose an original approach for discovering protein similarities based on their surface characteristics represented in terms of surface patterns. The approach starts from a detailed representation of the protein surfaces and determines a set of characteristic regions that defines a compact representation of the protein surface that is the input for an ad-hoc data mining technique used to find the frequent patterns. Tests, carried out on a benchmark dataset of molecules with suitably designed surface mutations, show that surface patterns can be used to correctly classify groups of similar proteins.