We experimentally learn structures of Bayesian networks classifying signals enabling genetic abnormality diagnosis. Structures learned based on the naive Bayesian classifier, expert knowledge or using the K2 algorithm are compared. Inferiority of the K2-based classifier has motivated an investigation of the algorithm initial ordering, search procedure and metric. Replacing the K2 search with hill-climbing search improves accuracy as does the inclusion of hidden variables into the structure. However, it is proved experimentally that this inferiority of the K2-based classifier is mainly due to the K2 metric soliciting structures having enhanced representability but limited classification accuracy.