Bayesian network (BN) [1] is a very simple and efficient classifier, and has a wide application in many fields.But its structure learning is quite difficult, especially for high dimensional data, such as liquid chromatography-mass spectrometry (LC-MS) data.Since the number of the possible structures grows exponentially with the number of variables.To process LC-MS data effectively, we applied an efficient BN structure learning method which constructs a two-level BN (BN-TwoL).BN-TwoL sets the class label C as the root of the network.The variables which are conditional independent of each other lie in different sub-trees of the root.