Natural selection operating in protein-coding genes is often studied by examining the ratio () of the rates of nonsynonymous to synonymous nucleotide substitution. The branch-site method (BSM) based on a likelihood ratio test is one of such tests to detect positive selection for a predetermined branch of a phylogenetic tree. However, because the number of nucleotide substitutions involved is often very small, we conducted a computer simulation to examine the reliability of BSM in comparison with the small-sample method (SSM) based on Fisher's exact test. The results indicate that BSM often generates false positives compared with SSM when the number of nucleotide substitutions is Ϸ80 or smaller. Because the value is also used for predicting positively selected sites, we examined the reliabilities of the site-prediction methods, using nucleotide sequence data for the dim-light and color vision genes in vertebrates. The results showed that the site-prediction methods have a low probability of identifying functional changes of amino acids experimentally determined and often falsely identify other sites where amino acid substitutions are unlikely to be important. This low rate of predictability occurs because most of the current statistical methods are designed to identify codon sites with high values, which may not have anything to do with functional changes. The codon sites showing functional changes generally do not show a high value. To understand adaptive evolution, some form of experimental confirmation is necessary.branch-site method ͉ small-sample method I n the current statistical methods of inferring positive selection using the value, it is assumed that Ͼ 1, ϭ 1, and Ͻ 1 represent positive, neutral, and negative selection, respectively (1). One of the statistical methods using this approach is the branch-site method (BSM) (2, 3). In this method, the branches of a phylogenetic tree are divided into a predetermined (foreground) branch and other (background) branches and codon sites are grouped into a few classes with different values (see Methods). The log likelihood (lnL) for the selection model used (modified model A) is then compared with that for the null model of no positive selection ( Յ 1), and the likelihood ratio test (LRT) is conducted to determine whether positive selection is operating in the foreground branch. This method has been widely used (e.g., 4-7), and one of the recent applications is Bakewell et al.'s (5) large-scale analysis of orthologous gene trios from humans, chimpanzees, and macaques. In this case, however, the numbers of synonymous (c S ) and nonsynonymous (c N ) substitutions per gene per branch were so small that the applicability of the large-sample theory of LRT is questionable.Another test that is applicable for this type of datasets is the small-sample method (SSM) using Fisher's exact test (8). In this method the ancestral nucleotide sequence at each interior node is inferred by the parsimony method, and c S and c N for the branch to be tested are counted by comparing the ...