Letter to the Editor 18S rRNA Sequences and Amniote Phylogeny: Reply to Marshall’

Marshall ( 1992) presents a reanalysis of some 18s rRNA data (Hedges et al. 1990) bearing on tetrapod relationships. Our phylogenetic analyses of these data for 26 tetrapods supported, among other groupings, a bird-mammal relationship. However, Marshall concluded that his reanalysis of the amniote data, using a weighted parsimony method (Williams and Fitch 1989, 1990)) supported a tree uniting birds and squamates (lizards and snakes). We believe that his results were due to a misapplication of the weighted parsimony method, which was designed for highly variable (noisy) datanot for highly conserved data. The controversy over amniote relationships involves not only the paleontological data and our 18s rRNA data but other morphological and molecular data sets (Hedges et al. 1990; Hedges and Maxson 199 1). The best estimate of amniote phylogeny will not hinge on any one data set unless such a set is very large; more sequence data will be needed to resolve this important question. However, the available 18s rRNA sequence data (Hedges et al. 1990, fig. Al; present paper, fig. 1) clearly support a birdmammal relationship. Weighted parsimony methods were developed by Sankoff and Cedergren ( 1983 ) and Williams and Fitch ( 1989, 1990)) to account for substitution biases in phylogenetic analysis. Marshall ( 1992) used the latter method to reanalyze our 18s rRNA data. The basic principle is that rare substitution types should be weighted more heavily than common substitution types because common substitutions are more likely to occur multiple times at the same site, thus obscuring phylogenetic information. The implementation of the weighting is done a posteriori by observing the frequency of the different substitution types in an initial tree and then weighting inversely to those frequencies. Each site also can be weighted inversely to the number of changes at that site in the initial topology. That there is substitution bias in the 18s rRNA data is not surprising, because it is present in most nucleotide sequence data sets. Mechanisms have been proposed for some types of biases (e.g., transition-transversion bias and codon bias), but the reason for the unequal frequencies of certain substitution types in the 18s rRNA data is presently unknown. One potential problem with weighted parsimony involves the basic assumption that rare substitution types are more reliable indicators of phylogenetic relationships. This concept is more useful when there is a high probability of multiple changes per site (multiple hits). Weighting rare changes more heavily in a highly conserved set of sequences (such as the 18s rRNA data) is unwarranted because all substitution types convey the same phylogenetic information (i.e., they are equally detectable) regardless of relative frequency. There is a wide “gray zone” where the relative information content of rare changes increases as the probability of multiple hits increases. It is unclear at what point (if any) an inverse weighing scheme proportionately compensates for the increasing noise in the data. Application of such a weighting scheme to a data set which has not reached this noise level constitutes a bias. Williams and Fitch ( 1989, 1990) intended their method to be used with noisy data sets. Although the presence of homoplasy in the 18s rRNA data indicates that