Résumé : We propose here to learn automata for the characterization of proteins families to overcome the limitations of the position-specific characterizations classically used in Pattern Discovery. We introduce a new heuristic approach learning non-deterministic automata based on selection and ordering of significantly similar fragments to be merged and on physico-chemical properties identification. Quality of the characterization of the major intrinsic protein (MIP) family is assessed by leave-one-out cross-validation for a large range of models specificity.