The article introduces an original problem of knowledge discovery from chemical reaction databases that is closely related to reaction clustering. The problem aims at identifying the subset of atoms and bonds that play an effective role in a given chemical reaction. The resulting characteristic reaction pattern describes the synthesis method underlying the reaction. Solution to this question is approached by a graph-mining optimization problem whose setting is new: given lower and upper bound graphs gl and gu, the search of best patterns in an interval of graphs consists in finding among graphs isomorphic to a subgraph of gu and containing a subgraph isomorphic to gl, best patterns that maximize a scoring function. Score values depend on the frequency of patterns in a set of examples. The article also studies accuracy and scalability of the method when applied to reaction databases.