Proteins perform very important functions within organisms. Predicting these functions is a major problem in biology. To address this issue, predictive models of functional families from the sequences of amino acids that form the proteins have been developed. The Dyliss team developed a machine learning algorithm, named Protomata-learner, that learns weighted automata representing these families and the possible disjunctions between members. New sequences can be compared to these models and assigned a score to predict their belonging to the family. Despite good results, the sequence weighting strategy and the null-models in Protomata are rather basic. During my internship, I investigated alternative sequence weighting strategies and null-models. Besides, the expressivity of Protomata leads to a great variability of scores and the choice of the classification threshold was left to the user. So, I proposed a normalization of the score, and a method to assess the significance of scores, to simplify the prediction. I implemented these new strategies and compared them on several data sets. Preliminary results show a good improvement of the prediction power of the computed models.