The paper is devoted to the detection of pair correlations in amino ac
id sequences using mutual information and correlation functions. Since
statistical dependences are relatively weak, finite sample correction
s turn out to be essential for a correct interpretation of the correla
tion measures. Statistical fluctuations are reduced by analyzing two l
arge sets of protein sequences. The calculation of correlation functio
ns requires the assignment of numbers to the 20 amino acids. This is d
one by using well-known property codes. Moreover, we look by extensive
random sampling for optimized codes which maximize the strength of co
rrelations. It turns out that the strongest correlations are related t
o hydrophobicity scales and a-helix propensities.