We suggest looking at their ingredients, one at a time, with everything else fixed, and considering how well they describe documented cases of functional divergence. We also stress the fact that EX 527 scoring functions, wittingly or not, often encompass an evolutionary model that cannot be applied across the board. While discriminants can be commonly found in catalytic sites of enzymes, they are more of an exception than a rule in a general case of functional divergence. When dealing with real-life data there are many additional practical problems that need to be resolved, and diverse sources of information that need to be collated. The estimation of the reliability of the alignment in the neighborhood of the residue of interest, treatment of gaps, unsupervised detection of orthologous groups mapping onto the structure, as well as detecting synergistic co-evolutionary events are all important issues, but downstream or complementary to the basic specialization scoring framework we propose to discuss here. In the following section, we lay out the framework for discussion of overlap and conservation measures. Therein we also outline the incorporation of residue exchangeability in the description, and show how these basic ingredients combine into various specialization scoring functions. In the Results section we take a look at several examples of specialization among families of paralogous proteins, and discuss where the responsible residues fall on the conservation/overlap grid. We consider the options available in building a scoring function at a heuristic, phylogeny independent level, and propose a strategy that allows us to move on from catalytic sites of enzymes to more general cases of protein functional divergence. First, we compare the performance of different specialization scoring schemes for cases where the difference between groups stems from the change in the nature of a small ligand binding site. This is the type of scenario where we are the most likely to encounter the “discriminant” types of positions: binding of a small ligand does not allow much freedom in the residue type choice. Different ligands, however, require different residue types. In such cases mutual information is expected to be a good measure for their detection. In one of the most thorough point-mutational studies of a protein we have up to date, Suckow and collaborators mutated almost all positions in E. coli lactose inhibitor from its wild type to 12 alternative amino acid types, and divided the resulting phenotypes into five distinct groups. The phenotype we are particularly interested in is the loss of inducer response – the trait that distinguishes LacI from its paralogous relatives, purine and galactose repressors. The size of this systematic study provided a precious set of true negatives, shown in the inset of the first panel, Fig. 2. In the main panel, the standardly used ROC curve, using residues not explicitly known to be involved in the specific function as the set of “negatives.” The behavior of different scoring methods indicates that while several of the specific residues behave as discriminants, the rest do not, and mutual information fails to locate them.