In contrast to previous analyses (1, 2), Gibbs et al. (3) did not use the avian lineage to root their phylogenetic trees, and they suggested that the position of the bird lineage depends on the substitution model used. However, using avian influenza to distinguish the human and swine lineages results in phylogenetic tree topologies (Fig. 1) that are almost identical for both putative recombinant regions proposed by Gibbs et al. (3). In particular, the 1918 strain is placed on the human lineage not just for the putative human region but, crucially, for the supposed swine region as well (with 96% bootstrap support). Thus, maximum likelihood trees incorporating the avian lineage provide a robust, informative, and necessary test of the recombination hypothesis, whereas the midpoint rooting employed by Gibbs et al. is affected by differences in the rate of molecular evolution between the HA1 and HA2 gene regions. The phylogenies (Fig. 1) show the same relative depth of the swine influenza clades for HA1 versus HA2, illustrating a nearly constant rate of evolution across the entire HA in this lineage. In contrast, for the human lineage, the HA2 region evolves at a considerably lower rate than HA1. Pairwise comparisons among the human lineage strains confirm that rate difference, with uncorrected distances roughly twice as high for the HA1 region as for the HA2 region.

Plots presented by Gibbs et al. [figures 1A and B in (3)] give the strong impression that the 1918 strain is more similar to the swine lineage in the HA1 region than across the remainder of the HA gene. The 1918 sequence, however, actually exhibits about 11% uncorrected evolutionary distance to the swine reference strain across both regions (Fig. 2B). That fact is obscured because Gibbs et al. plotted the percent nucleotide identity at variable sites only, and the rate change in the human lineage will affect that proportion (Fig. 2A). There are considerably more variable sites in the HA1 region, where the human lineage evolves at a relatively high rate, than in the HA2 region, where the human lineage evolves at a relatively low rate (Table 1). Although there are about the same number of differences between the 1918 strain and the swine-lineage strain in each region, the similarity at variable sites is higher in the HA1 region than in the HA2 region. This effect will give the misleading impression of recombination, particularly when a relatively ancient strain that falls near the root of the two lineages is compared with reference strains from each lineage. The standard similarity plot, constructed using all sites (Fig. 2B), reveals that the 1918 HA sequence exhibits no greater affinity with the swine-lineage sequence in HA1 compared with HA2. The distance between the 1918 strain and the human-lineage reference strain, however, is about 17% in HA1 and 8% in HA2, a difference that is accounted for by the human-lineage rate bias evident in the phylogenies.

Table 1. Variable sites in the recombinant regions in HA1 and HA2 identified by Gibbs et al. (2). V, number of variable sites; Dhuman, number of sites at which 1918 strain differs from human strain; Dswine, number of sites at which 1918 strain differs from swine strain. V Dhuman Dhuman/V (%) Dswine Dswine/V (%) Globular domain region (sites 310-870) 142 94 66.2 61 43.0 Stalk domain region (sites 1070-1650) 95 45 47.4 64 67.4

Finally, we simulated sequence data on a clonal tree, but incorporated the observed rate difference in the human lineage of HA1. Split-decomposition graphs (4) obtained from the simulated data were very similar to the graph obtained from the real data (Fig. 3), an indication that the rate difference across the different regions of the human-lineage HA gene--and not recombination--was the cause of the network pattern in the split-decomposition graph for the real data.

We conclude that there is no evidence that the HA gene of the 1918 Spanish flu virus had a recombinant origin. The finding of Gibbs et al. was the result of a localized difference in the rate of molecular evolution along the human-lineage HA gene. In view of the exceptional virulence of the 1918 epidemic, further investigations into the processes underlying the patterns of influenza sequence variability are required.

Michael Worobey
Andrew Rambaut
Oliver G. Pybus
David L. Robertson
Department of Zoology
University of Oxford
South Parks Road
Oxford, OX1 3PS, UK
E-mail: david.robertson{at}zoo.ox.ac.uk

REFERENCES AND NOTES

Response: Worobey et al. point to evidence of constraining selection on the part of the HA gene that encodes the stem of the protein. This selection differentially affected the human but not the swine lineage of H1 influenzas. It had not been reported before and we had not recognized it, even though it was evident in our results. It affects a large part of the gene, which suggests that structural changes in the stem are selected against. Worobey et al. propose that the evidence of recombination we reported in the HA gene of the 1918 influenza (1) results from this differential selection.

This is an interesting idea, but they have not yet proven it. A phylogenetic analysis is not a "necessary test" for evidence of recombination (2). We doubt that the avian lineage may be used to root the tree of the mammalian lineages, because, like others (3), we found that the relative positions of the 1918 sequence and the avian lineage vary in trees, depending on what nucleotide sites or methods were used. We also found evidence that the HA sequences from birds and mammals are evolving quite differently, and we know of no way to prove which phylogenetic method or model is appropriate for this complex, incoherent data set (2). Current implementations of the maximum likelihood method used by Worobey et al. fit a single model to an entire dataset; thus, the positions of nodes (like the 1918 HA sequence) that link, or fall between, lineages that have evolved in different modes must be less than certain.

The degree of conservation and the strength of an evolutionary signal usually vary along a gene sequence, and it is a change in the relative strengths of several signals that is generally considered to indicate recombination (4). We used sister-scanning (5) to judge the relative support for the pairwise relatedness of four aligned sequences in a succession of windows sampled from along an alignment. Variations in the relationships between the swine and human lineage sequences were expected. By ignoring invariant sites, we eliminated one of the major sources of sequence variation that does not contribute anything to the relative signals. Relatedness can be expressed in various ways, but the values for each window are calculated quite independently of the others. Thus, our finding that, in a series of windows, the HA1 region of the 1918 HA gene sequence is significantly more closely related to the same region of swine lineage HA genes than to human lineage genes is independent of any features of the HA2 regions of the same sequences. There is no obvious reason to suspect that calculations showing the 1918 HA1 region to be more closely related to swine HA1 regions than human HA1 sequences were biased by differences in evolutionary rates.

The 1918 HA gene is most closely related to the HA gene of the oldest "classical swine" isolate (3), and we also found similar, although not identical, evidence of recombination in the HA gene sequence of that isolate. The proposal of Worobey et al. might explain the patterns found in the 1918 sequence, but it does not explain the patterns found in the HA gene of the classical swine isolate, which, for various reasons (1), we consider to be a member of the same recombinant lineage.

Mark J. Gibbs
John S. Armstrong
Adrian J. Gibbs
School of Botany and Zoology
The Australian National University
Canberra ACT 0200, Australia
E-mail: Mark.Gibbs{at}anu.edu.au

REFERENCES