We generalize the classification noise model of PAC learning introduced by  Angluin and Laird (1988) to a model, where the noise can depend on the input vector. To be able to devise efficient learning algorithms,  we utilize statistical queries due to Kearns (1998). Haussler's covering method is our vehicle for demonstrating how these ideas work in theory and practice. Having modifying its cover phase,   we apply it to the problem of finding a powerful discriminator between protein-protein interfaces and random sets of pairs of protein surface residues.  The hypothesis learned is convincing with respect  to the classification rate as well as  to its biological interpretation.