Figure 2 shows all the segregating sites for the pre-/early-vaccination era sequences with positions marked by base pair, and whether a substitution is synonymous or non-synonymous or both (for multiple hits and complex codons). It is immediately clear that many incompatibilities (by the four-gamete test 27) exist, but also that apparent groups of physically close sites are incompatible with other such groups. We identified by eye five such groups of informative sites and marked them with different colours. Comparing these five groups by the four gamete test demonstrates that some pairs of blocks are incompatible with the same evolutionary tree (Fig. 2b). Even though the identification of these groups is subjective, the presence of such groups suggests either that they have evolved on different evolutionary trees (i.e. recombination) or, alternatively, each of the sites for a given group has mutated at least twice, i.e. by parallel mutations, at approximately the same time (since there is strong linkage disequilibrium within all groups except block 1). Note that the incompatibilities within e.g. block 1 do not weaken this conclusion. A similar table was prepared for the post-vaccination era data set. Here we were not able to visually identify groups of sites failing to satisfy the four-gamete test (results not shown), and the blocks identified in Figure 2 are not incompatible with one another in the post-vaccination data set even though it contains all of the segregating sites. Thus, if the patterns observed in Table 1 were due to convergent evolution of groups of functionally important sites in different lineages, such convergent evolution does not play an important a role in present-day evolution of measles virus. The proportion of pairs of informative sites which is incompatible by the four gamete test is 43% in the pre-/early-vaccination data set as compared to only 23% in the post-vaccination data set even though the variability in general is lower among the pre-/early-vaccination sequences.

One way to distinguish whether incompatibilities are caused by parallel mutations (i.e. true homoplasies) or by recombination is to correlate linkage disequilibrium (LD) between sets of informative sites to distance. Recombination causes a decrease in linkage disequilibrium and more recombination is expected between sites that are further apart from each other. The block structure of close sites is visualised in a different way in Figure 3, which shows all the significant cases of linkage disequilibrium. Visual inspection of the Figure suggests that physically close sites appear to be more likely to be in strong LD than sites far apart. This observation is tested more formally by correlating LD with distance (Table 2 and Figure 4). Table 2 shows the correlation of two commonly used measures of LD for both data sets. A significant negative correlation is observed by both measures of linkage disequilibrium in the pre-/early-vaccination data set, whereas the post-vaccination data set does not show any significant decay of linkage disequilibrium with distance. The decrease in linkage disequilibrium with distance for the pre-/early-vaccination sequences is shown graphically in Figure 4.

An analysis of expected number of homoplasies caused by parallel mutations and the number of apparent homoplasies in the most parsimonious tree was performed on the pre-/early-vaccination data set [see 3235]. The basic idea is to investigate whether the number of apparent homoplasies in the most parsimonious tree is likely to be due to parallel (recurrent) mutations. If this is not the case, it is a strong indication of recombination since recombination easily creates "incompatible pairs of sites", i.e. pairs of sites that do not fit into the same phylogenetic tree, and thus will appear as homoplasies if assuming no recombination and a single phylogenetic tree. There are 800 sites. Of these, 267 are third position sites, and among these we observe 20 transition mutations. If we assume equal mutation rates at all third position sites, then the number of transition mutations we expect have happened while observing 20 different ones is 20.8+/-0.9 (calculated as the sum of geometric distributions). In the most parsimonious tree found using PAUP*, there are 26 transition mutations. In other words, given the observed number of mutation events of the transition type, we expect 0.8+/-0.9 parallel mutations/homoplasies under the assumption of no recombination (and no codon usage bias). We observe, however, 6 parallel mutations in the tree. This suggests that some of the apparent homoplasies resulted from recombination rather than from recurrent mutations. We also calculated the effective sites number 35 from the observed codon usage bias. However, since the codon usage bias is low, the effect is minor, and only 1.0+/-1.0 homoplasies are expected under this model, again significantly smaller than the six apparent homoplasies. A large amount of mutation rate heterogeneity at the synonymous sites offers an alternative explanation of the excess homoplasies. While this explanation cannot be ruled out, analysis of the post-vaccination data set does not support large rate heterogeneity at silent sites in the evolution of present-day measles virus, since the number of homoplasies in this data set can be explained by recurrent mutations (results not shown).

Given these different lines of evidence for recombination, it is of interest to try to estimate the rate of recombination needed to explain the data. The most appropriate method is the finite site, composite likelihood approach implemented in LDhat 31. The result (Table 2) is that a significantly positive recombination rate is found in the pre-/early-vaccination sequences, much reflecting the results of the similar R² test. An estimated rate of ρ = 15.8 corresponds to that an expected 45 recombination events have happened in the ancestral history of the 18 pre-/early-vaccination sequences. The estimated recombination rate appears smaller than rates reported for HIV and other viruses 31. The estimated recombination rate in the post-vaccination era data sets also shows a positive rate of ρ, but it not significantly different from zero by the permutation test (Table 2). The lack of evidence of recombination in present day sequences of MV strains is consistent with what was also observed by 31.

In conclusion, the analysis of pre-/early-vaccination era MV sequences shows evidence of recombination at rates important to the evolution of MV. The five different tests of recombination should not be considered independent tests and some of the results might be explained by alternative mechanisms such as convergent evolution of functionally important sites and rate heterogeneity of synonymous variation. However, all tests agree and provide evidence of recombination both through an excess of apparent homoplasies compared to the expected frequency of parallel mutations, and through a decrease in LD with distance, which is difficult to explain by any other hypothesis than recombination. Furthermore, it is difficult to imagine a mechanism other than recombination by which apparent homoplasies could occur pairwise or in triplets in distant parts of the sequence considering also that such patterns are not seen in the post-vaccination data set.

The evidence of recombination among pre-/early-vaccination era MV strains and the lack of detection of recombination among post-vaccination era MV strains are consistent with the shift in epidemiology from a situation of co-circulation of strains in populations to a bottleneck situation with incidental introduction of a single strain to a susceptible sub-population of a geographical region. Given that intergenomic recombination and co-infection of individuals are common phenomena of MV it might be assumed that the lineages surviving till today did emerge from a pool of recombining pre-vaccination era strains. A lower level or absence of recombination due to changed epidemiology since then has erased our ability to detect recombination in a global sample of present day MV despite its high level of variability. It is possible that the Danish pre-/early-vaccination era strains are representatives of the very pool in which recombination took place while present-day MV strains are representatives of temporally and/or geographically separated lineages. Analysing informative sites in other parts of the MV genome of present-day lineages render it unlikely that these lineages could have derived from a clonal population structure of a global pool of MV strains (L. S. Christensen, unpublished data).

The template for replication in members of the Paramyxovirinae is a nucleocapsid complex in which each nucleocapsid monomer (N) is predicted to be associated by hydrophobic bonds with 6 nucleotides in such a way as to resist non-ionic detergent and high salt and to protect the RNA from RNase digestion 36. This tight association of RNA and protein has been dubbed "the rule of six" and excludes the intracellular presence of naked viral RNA molecules. It raises the question of the mechanism of RNA polymerase template recognition and has provided an explanation of the hypothesis that recombination possibly cannot occur in this group of viruses. Our data suggests that a mechanism of partial unwinding of the nucleocapsid structure exists to allow homologous intergenomic recombination or RNA polymerase template shift during replication.

The Danish pre-/early-vaccination sequences consist of an 800 base-pair region (nt. 659 to 1458) of the hemagglutinin coding region of 18 strains collected in Denmark, Greenland and the Faeroe Islands between 1965 and 1983 (GenBank accession numbers AJ417850-AJ417867) 15. Post-vaccination sequences of 40 strains, representative of the 22 phylogenetic clusters identified 171819, were trimmed to match the region of the pre-/early-vaccination era sequences. GenBank accession numbers of 33 of these sequences can be found in 15. The remaining 7 HA sequences that complete the list of globally circulating genome types, identified at present, are those of strains Mvi/Gambia/93 (Type B3, Acc. No AF484955) 20, MVi/Alberta.CAN/20.00/1 (Type D7, Acc. No. AF410986) (G.A. Tipples et al., submitted 14-Aug-2001), MVi/Montreal.CAN/19.98 (Type D8, Acc. No. AF410985) (G.A. Tipples et al., sumitted 14-Aug-2001), MVi/Vic.AU/12.99 (Type D9, Acc No. AY127853) 21, (Type G2, Acc. No. AF243851) 22, MVi/Gresik.INO/18.02 (Type G3, Acc. No. AY184218) (P.A. Rota and S.L. Liffick, submitted 19-Nov-2002), and China94-1 (Type H2, Acc. No. AF045203) 23.

The alignment of the 18 pre-/early-vaccination and 40 post-vaccination sequences does not contain any gaps. The computer program DnaSP 3.99 24 was used for estimation of standard parameters of population genetics. Segregating sites were classified as synonymous or non-synonymous, except for complex codons where a site may be classified as both. Phylogenetic trees from the post-vaccination data set and the two data sets combined were built using MEGA 2.1 25 and the minimum evolution criterion. The HKY substitution model 26 was assumed. Bootstrap values were estimated from 2000 re-samples.

Recombination was examined using five different, but complementary, approaches. (i) A graphical method by which sets of sites in strong linkage disequilibrium (LD) were visually identified and marked with different colourings (Fig. 2). It was then investigated whether different sets of blocks were incompatible by the four-gamete test 27. If blocks are incompatible, one will infer either recombination or recurrent mutation of the sites in a block at about the same time. (ii) A plot of significant pairwise linkage disequilibria was constructed using DnaSP. (iii) Decay of linkage disequilibrium with distance (measured as either the squared correlation coefficient R² or the standardized measure D') was investigated following 28 and estimated using the R2-program 29. Analysis was restricted to informative sites. (iv) Estimation of the scaled population recombination rate ρ by a composite maximum likelihood approach 30, using the LDhat program of 31. This program also allows a test of the null hypothesis of no recombination (ρ = 0) by a permutation test. This analysis was also restricted to informative sites. (v) Comparison of the expected number of homoplasies to the number of apparent homoplasies in the most parsimonious phylogenetic tree, closely following the approach of 32, using PAUP* 32.