In the last years, pattern recognition has been applied to several fields for solving multiple problems in science and technology as for example in protein prediction. This methodology can be useful for prediction of activity of biological molecules, e.g. for determination of antimicrobial activity of synthetic and natural peptides. In this work, we evaluate the performance of different physico-chemical properties of peptides (descriptors groups) in the presence of imbalanced data sets, when facing the task of detecting whether a peptide has antimicrobial activity. We evaluate undersampling and class weighting techniques to deal with the class imbalance with different classification methods and descriptor groups. Our classification model showed an estimated precision of 96% showing that descriptors used to codify the amino acid sequences contain enough information to correlate the peptides sequences with their antimicrobial activity by means of learning machines. Moreover, we show how certain descriptor groups (pseudoaminoacid composition type I) work better with imbalanced datasets while others (dipeptide composition) work better with balanced ones.