Models that explicitly account for the effect of selection on new mutations have been proposed to account for “codon bias” or the excess of “preferred” codons that results from selection for translational efficiency and/or accuracy. In principle, such models can be applied to any mutation that results in a preferred allele, but in most cases, the fitness effect of a specific mutation cannot be predicted. Here we show that it is possible to assign preferred and unpreferred states to amino acid changing mutations that occur in protein domains. We propose that mutations that lead to more common amino acids (at a given position in a domain) can be considered “preferred alleles” just as are synonymous mutations leading to codons for more abundant tRNAs. We use genome-scale polymorphism data to show that alleles for preferred amino acids in protein domains occur at higher frequencies in the population, as has been shown for preferred codons. We show that this effect is quantitative, such that there is a correlation between the shift in frequency of preferred alleles and the predicted fitness effect. As expected, we also observe a reduction in the numbers of polymorphisms and substitutions at more important positions in domains, consistent with stronger selection at those positions. We examine the derived allele frequency distribution and polymorphism to divergence ratios of preferred and unpreferred differences and find evidence for both negative and positive selections acting to maintain protein domains in the human population. Finally, we analyze a model for selection on amino acid preferences in protein domains and find that it is consistent with the quantitative effects that we observe.