Among the 38 whole-genome HCV sequences representing 18 confirmed subtypes as summarized in Table 1, the most general substitution model, the general time reversible model (GTR, also known as REV) with a discrete gamma approximation for rate heterogeneity, was consistently supported as superior among the twelve nucleotide substitution models evaluated (not shown). Models adjusted for rate heterogeneity consistently fit the data better than models that assume a fixed evolutionary rate across sites (not shown). Substitution models with fewer parameters or an assumption of equal base compositions performed significantly worse than GTR, regardless of whether or not the sequences analyzed contained protein-coding regions. Adding a parameter for the estimated proportion of invariant sites significantly improved the substitution model, yielding parameters as shown in Table 2. The same model was selected when the AIC was adjusted to compensate for a low ratio of sample data to parameters (not shown). Thus, GTR with a gamma distribution of evolutionary rates per site and accommodation of invariant sites (GTR+Γ+I) is the best substitution model for HCV variation among those considered, and was used for maximum-likelihood phylogeny inference.

The 5' UTR is represented by the smallest number of aligned nucleotide sites (300 nt; the 5' most 42 nt were excluded from analysis because of extensive gaps throughout the available sequence data), followed by the Okamoto region of NS5B (329 nt), then the polyprotein (9177 nt), and the whole genome (9791 nt, Table 2). The proportion of invariant nucleotide sites for the 5' UTR is 2/3, much lower than for the protein-coding regions, for which less than 1/3 of sites do not vary (Table 2). The 5' UTR is known to be less variable than protein-coding regions of HCV 361112.

Tree topologies from the entire HCV genome and the polyprotein are identical (Figs. 1a, b and 2a, b). The tree from the Okamoto region of NS5B resembles trees from the whole genome and the polyprotein, except for rearrangements in the ordering of deeply rooted branches (Figs. 1d and 2d). Trees from sequences that include protein-coding regions clearly group subtypes from the same genotype into clades, while the tree from the non-coding terminus conflates subtypes of genotypes 1 and 6 with subtypes 4a and 5a, and subtypes of genotypes 1 and 6 cannot be distinguished (Figs. 1c and 2c). Thus, the phylogenetic trees of the 5' UTR are less able to group subtypes from the same genotype together into clades than trees from protein-coding sequences (Figs. 1 and 2), regardless of the method used for phylogenetic inference. Parsimony analysis yields comparable results, with similar trees for the whole genome, polyprotein, and the Okamoto region of NS5B, while the tree from the 5' UTR contains a basal polytomy that does not resolve genotypes 1,4, 5, or 6 (not shown).

Log-likelihood scores and SH-test results for alternative trees are summarized in Table 3. All tests yield the same outcomes, regardless of whether or not RELL optimization was used. Comparisons of alternative trees with the 5' UTR data fail to reject the null hypothesis of no difference in likelihoods (P > α; see Methods). Comparisons among alternative trees with data from the Okamoto region of NS5B indicate that the 5' UTR tree has a significantly different likelihood (P < 0.0001) than trees obtained from NS5B, polyprotein, or whole-genome data, which are statistically indistinguishable (P > α). Comparing parsimony trees from 300-nt windows in NS5B with trees from the 5' UTR via the incongruence length difference test 13, which uses the difference in tree lengths as a test statistic, rather than the likelihood difference, yielded the same pattern of significant differences (not shown).

Increasing window sizes represent the CI as an increasingly smooth function, as more nucleotides better approximate the whole-genome phylogeny than fewer nucleotides. However, increasing window size yields poorer resolution in the 5' UTR (Fig. 3a) because fewer windows are able to represent this region. Contrary to expectations, the rescaled homoplasy index is not constant. Despite large fluctuations within the 5' UTR, the rescaled homoplasy index is generally greater in the 5' UTR than in other regions of the HCV genome and particularly NS5B (Fig. 3b). After correcting for the substitution rate in this manner, the consistency of sites with the whole-genome phylogeny is lower in the 5' UTR than in NS5B.

We used multiple methods for phylogenetic inference, including neighbor joining (NJ), maximum parsimony (MP), and maximum likelihood (ML) 910. This was done to evaluate whether the inferential technique has an influence on the ability of the resulting phylogenies to resolve subtypes into clades. We used PAUP*, version 4.0b10 40 for phylogenetic inference. Neighbor-joining trees were constructed with the F84 distance metric 41 and the BioNJ algorithm 42. For parsimony analyses, uninformative invariant characters were excluded and gaps were treated as a fifth character state.

To select an appropriate nucleotide substitution model, we used FindModel, an independ-ent, online implementation of ModelTest 43. This approach uses an information-based goodness-of-fit criterion, in the sense that the best model minimizes the quantity of bits required to encode both the model and the model-encoded data for electronic transmission 444546. Such an approach includes a penalty term for the number of parameters, and thus facilitates comparing models with varied numbers of parameters 44. The fit of each model to the data was evaluated both with and without a four-category discrete approximation to a gamma distribution of substitution rates per site. Because FindModel does not test models with invariant sites, we also used ModelTest (version 3.6) to evaluate nucleotide substitution models with invariant sites 43. Akaike's information criterion (AIC) was used to quantify the suitability of alternative models having varied numbers of parameters to fit the data 47.

To evaluate the significance of differences in ML phylogenies obtained from different regions of the HCV genome, we used the Shimodaira-Hasegawa (SH) test 48 as implemented in PAUP*, version 4.0b10 40. The null hypothesis of the SH test is that none of the trees evaluated has a likelihood that differs significantly from any other. Rejecting the null hypothesis indicates a significant difference in likelihood scores, and thus in tree topologies 49.

For a pair of trees defined a priori, the SH test computes the difference in their likelihoods (Δ). This difference is compared with the null distribution of likelihood scores, obtained by building trees from character data generated by iterative bootstrap resampling with replacement of the nucleotide sites. A computationally efficient optimization (RELL) may be applied, which simply adds together per-site likelihoods over the resampled sites. Otherwise, the tree parameters are optimized on the resampled data (FULL). The resampled likelihood differences are denoted Δ′i MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacuqHuoargaqbamaaBaaaleaacqWGPbqAaeqaaaaa@2FA5@, where i indexes the replicate, and they are subsequently transformed by subtracting the mean resampled difference <Δ'>, a procedure called centering. The original difference in likelihoods is compared with the null distribution in a one-tailed, non-parametric manner, whereby the rank of Δ is evaluated against the centered, sorted Δ' distribution. If the rank of Δ is found to lie outside the interval of the null distribution between 0 and the (1-α) × 100 percentile, the difference in likelihoods is significant with (1-α) × 100% confidence, and the null hypothesis is rejected in favor of the alternative. (The acceptable type I, or false positive, error rate per test is denoted α.)

Here the tree topologies are ML phylogenies that represent different regions of the HCV genome. The reference alignment of 38 HCV whole-genome sequences representing 18 confirmed subtypes (Table 1) was obtained from the LANL HCV database 50. We conducted SH tests with data from the 5' UTR, the Okamoto region of NS5B, and whole genome. Topologies were paired such that the ML tree Tx∗ MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGubavdaqhaaWcbaGaemiEaGhabaGaey4fIOcaaaaa@3072@ inferred from the data of region x (either the 5' UTR or Okamoto region) was compared with the ML tree Ty∗ MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGubavdaqhaaWcbaGaemyEaKhabaGaey4fIOcaaaaa@3074@ from data of region y representing each other region (either 5' UTR, Okamoto region, polypeptide, or whole genome, provided y ≠ x), yielding the likelihood difference Δ ≡ L_x(Tx∗ MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGubavdaqhaaWcbaGaemiEaGhabaGaey4fIOcaaaaa@3072@) - L_x(Ty∗ MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGubavdaqhaaWcbaGaemyEaKhabaGaey4fIOcaaaaa@3074@), where L_x(Ty∗ MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGubavdaqhaaWcbaGaemyEaKhabaGaey4fIOcaaaaa@3074@) is the likelihood of the ML tree from region y evaluated with data from region x. We randomly resampled 10,000 replicate data sets for each pair of trees and compared the original difference in likelihoods with the null distribution that resulted. The type I error rate was reduced to accommodate six hypothesis tests (α = 0.05/6 = 0.00833). This reduction preserves the experiment-wide false-positive rate by making each comparison more stringent.

To understand better phylogenetic inconsistencies over the HCV genome, we computed the character consistency index (CI) for each site in PAUP with the whole-genome phylogeny, and summarized CI with a moving-window (running) average over 100, 300, and 500 nt. The 100 nt window size was used subsequently because it allows for clear visualization of the 342 nucleotides that constitute the 5' UTR. Because the consistency and homoplasy indices (HI) are complementary (CI+HI = 1), character consistency is high when homoplasy is low, and vice versa. Thus, we expect lower homoplasy to result from fewer informative sites. Further, homoplasy decreases rapidly with decreasing substitution rates. To control for variation in the number of informative sites across the genome, we rescaled the homoplasy index against the square of the proportion of informative sites in the window region. This was done because, in the limit of short branch lengths, the number of informative sites should be proportional to the substitution rate r, while the number of homoplasies should be proportional to r². The result was subsequently normalized against the maximum, to facilitate comparison with the proportion of informative sites. As a result, if all parts of the HCV genome are equally informative, one can expect the rescaled homoplasy index to be roughly constant over the viral genome.