We describe a set of tests designed to check the ability of the new "membrane score" method (see the first paper of this series) to assess the packing quality of transmembrane (TM) alpha-helical domains in proteins. The following issues were addressed: (1) Whether there is a relation between the score (S(mem)) of a model and its closeness to the "nativelike" conformation? (2) Is it possible to recognize a correct model among misfolded and erroneous ones? (3) To what extent the score of a homology-built model is sensitive to errors in sequence alignment? To answer the first question, two test cases were considered: (i) Several models of bovine aquaporin-1 (target protein) were built on the structural templates provided by its homologs with known X-ray structure. (ii) Side chains in the spatial models of visual rhodopsin and cytochrome c oxidase were rebuilt based on the backbone scaffolds taken from their crystal structures, and the resulting models were iteratively fitted into the full-atom X-ray conformations. It was shown that the higher the S(mem) value of a model is, the lower its root-mean-square deviation is from the "correct" (crystal) structure of a target. Furthermore, the "membrane score" method successfully identifies the rhodopsin crystal structure in an ensemble of "rotamer-type" decoys, thus providing the way to optimize mutual orientations of alpha-helices in models of TM domains. Finally, being applied to a set of homology models of rhodopsin built on its crystal structure with systematically shifted alignment, the approach demonstrates a prominent ability to detect alignment errors. We therefore assume that the "membrane score" method will be helpful in optimization of in silico models of TM domains in proteins, especially those in GPCRs.