Why genes overlap in viruses

1Gene overlaps, which we define here as having nucleotides coding for more than one protein by being read in multiple reading frames, are a common feature of viruses. Proteins created by gene overlaps (sometimes called ‘overprinting’) are typically accessory proteins that play a role in viral pathogenicity or spread. These overlaps are typically assumed to be a form of genome compression, allowing the virus to increase its repertoire of proteins without increasing its genome length.

2Over the past several decades, many authors have suggested explanations for why gene overlap has arisen and become so common in viruses. In this study, we compare the amount of gene overlap across all known virus species to investigate the plausibility of these explanations. We find that the evidence is most consistent with the main effect being a physical constraint by the capsid (the protein capsule, into which the genome is packaged for transmission between host cells).

3RNA viruses make up approximately one-half of the 2000 known species of virus. They have an extremely high mutation rate and several authors have suggested that this could explain the evolution of gene overlap. The causality might be indirect: because most mutations are harmful, the high mutation rate will limit genome length, and thus new genes or gene regions must come from overlapping. Alternatively, several studies have shown how gene overlap in theory might mitigate the detrimental effects of mutation: (i) a numerical simulation shows a benefit of overlapping except in the case of synergistic epistasis between fitness traits, which is rarely observed in RNA viruses, (ii) an analytical model quantifies the increasing harm of mutations in overlapping genomes in terms of the information cost and comes to the same conclusion: a fitness advantage of overlap when mutation rate is high, and (iii) overlapping genes can be seen as one of many ‘antiredundant’ mechanisms that may lead, in the case of mutation, to the damage of distinct functions simultaneously, and one which facilitates the removal of mutant genomes from the population (purging; ). The authors show, through numerical simulations, that this kind of mechanism is likely to evolve in large populations, such as viruses. A more detailed study concludes that genome compression can increase the stability of the wild-type both by reducing mutation incidence (the advantage discussed above) and by reducing sequence redundancy.

4This general argument predicts that DNA viruses, which make up the other half of the known virus species and which tend to have a lower mutation rate, will have less gene overlap.

5Gene overlap might have evolved if genome length is physically limited by the size of the capsid. This was suggested over 30 years ago and has been invoked since to explain individual gene overlaps. Some observations are consistent with capsid size constraining genome length: in most viruses studied, it is not possible to package an artificially enlarged genome, and many studies on different virus groups have found virus genome length to be positively correlated with capsid size. Furthermore, the hypothesis is testable because, as we describe below, some capsid types might be expected to constrain genome length more than others.

6Many viruses have capsids that are icosahedral (20 sided), varying in the number of protein units (capsomers) that form each side. In such viruses, an increase in capsid size is generally achieved through increases in the number rather than the size of these capsomers. These increases in capsomer number are in discrete steps following a geometric pattern represented by the so-called T (Triangulation) number series, which appears to be thermodynamically determined. As the T values increase, the differences in volume (as a percentage) between adjacent T numbers become smaller, with the product of the capsid diameter and the reciprocal of the square root of T remaining constant. The actual pattern of historical transitions between different T numbers is unknown and probably determined by the type of fold found in the capsomer. We discuss this in the electronic supplementary material. In figure 1, we illustrate the general principle using some DNA viruses, chosen because both their T number is known, and because of capsomer fold similarity, we can infer the likely next highest possible T number.

7The critical point is that a small virus with an icosahedral capsid, unlike a large virus with an icosahedral capsid, cannot physically make relatively minor adjustments to the size of its capsid by increasing its T number. We speculate that such viruses are therefore more likely to acquire novel gene function via overlap. The situation is different among viruses with non-icosahedral capsids because in theory adjustments in capsid size, and hence genome length, are as easy to make for small viruses as large ones, e.g. individual capsomers can simply be added onto the end of a helical capsid, as shown for M13 phage and Tobacco Mosaic Virus. We therefore predict that the negative relationship between gene overlap and genome length known at least for RNA viruses will be stronger among viruses with icosahedral capsids than among viruses with non-icosahedral capsids.

8The negative relationship between overlap proportion and genome length mentioned above could merely be an artefact of a relationship between gene length and genome length, and hence not require any biological explanation. The length of RNA virus genes with replicase functions increases with genome length; indeed, there may be a general tendency for genes to become larger in taxa with larger genomes, e.g. eukaryotes tend to have longer genes than prokaryotes. It appears that most gene overlaps in RNA viruses started from two originally contiguous genes, so if mean gene length increases with genome length, then there will be fewer opportunities for overlaps to evolve in a given length of nucleotides. We investigate the importance of this relationship.

9Some gene overlaps may have evolved to couple gene expression rather than to compress the genome. This is thought to be common in bacteria, and there are a few possible examples in viruses, e.g. in the RNA phage MS2 the start of the lysis gene overlaps with the end of the coat gene; translation of the lysis gene requires coat protein synthesis termination followed by reinitiation, the frequency of which may be a mechanism to regulate relative protein levels. Suspected cases of gene regulation typically involve short terminal overlaps, so we explored the effect of excluding from our analyses all overlaps of length less than 60 bases, which is approximately the minimum size of a functional protein.

10We cannot use measures of overlap in individual species as data for statistical tests because many species will have the same overlap between homologous genes. Such overlaps will often have a common origin and hence individual species do not represent independent data. However, we can use virus families as independent data because there is very little homology across families and thus most gene overlaps are likely to have been independently acquired. For example, in RNA viruses the only putative homology across all families is a small region of their replicase. We therefore use family means throughout. We follow the taxonomy of the NCBI Virus Genome website (‘GenBank’; currently at http://www.ncbi.nlm.nih.gov/genomes/GenomesHome.cgi?taxid=10239), and our analysis was based on a downloaded RefSeq flat file of July 2008. Unassigned genera and species were treated as additional independent data points. We follow NCBI's placement of Hepadnaviridae and Caulimoviridae among the RNA viruses despite their mature virion containing DNA. This is sensible owing to their possession of reverse transcriptase, which clearly allies them to reverse-transcribing RNA viruses.

11We measured gene overlap only as the overlap between the reading frames of different genes, i.e. we ignored regulatory regions (which are often unknown) or overlaps in the same frame. We also only included genomes that are classified by NCBI as either reviewed or validated (see below for exceptions). In the case of segmented viruses, we only included species where all segments were thus classified. The taxa included and excluded in our analysis, along with mean overlap proportion and genome length, are listed in the electronic supplementary material, table S1. We also added the recently discovered overlaps in Potyviridae, Reoviridae and Sequiviridae, and the overlap in Schizochytrium single-stranded RNA virus. We were thus able to calculate means of gene overlap for 36 RNA and 26 DNA virus families (or unassigned genera or species). Exclusion of the less well-annotated species is important, e.g. their inclusion obscures the underlying relationship between overlap proportion and genome length among double-stranded (ds) DNA viruses (electronic supplementary material, figure S1). Excluded data represent seven families plus 10 unassigned genera or species of DNA virus, and 10 families plus 21 unassigned genera or species of RNA virus.

12For analysing the trends in overlap proportion between families, we excluded families that have no gene overlap and then used natural logarithms of the overlap proportion and genome length to obtain approximate linear relationships. This exclusion is necessary for logarithmic transformation and should not affect our findings: only one DNA virus family, the Nanoviridae, has no overlap (as discussed above, we are only considering families represented by at least one validated or reviewed RefSeq entry). This family is unusual in having multiple small circular segments each coding for a single protein. Eight RNA virus families are also excluded because they appear to have no gene overlap. Some of these probably do have some gene overlap: some Bunyaviridae provisional RefSeq entries have gene overlap, and an overlapping gene has been reported in a member of the Picornaviridae—Theiler's murine encephalomyelitis virus (Theilovirus; ) and in a member of the Dicistroviridae. In order to compare the amount of gene overlap in RNA and DNA viruses, we calculated the mean of the family means, both including and excluding the small number of families that lacked overlap.

13We included all the estimated mutation rates that we could find in the literature (electronic supplementary material, table S3). Rates are typically expressed as the number of substitutions per base per round of genome copying. A few studies give the rate per round of cell infection, which will be higher but not misleadingly so, especially given the likely error margins on these values. Despite an extensive search, we could find estimates for only six DNA viruses, only two of which have genome lengths within the range of RNA viruses.

14Using a standard reference work, we classified virus families (and unassigned genera and species) as having either icosahedral or flexible capsid types. We treat as flexible all non-icosahedral types, e.g. capsids described as spherical, filamentous, helical or rod-shaped, or where there is no well-defined capsid (electronic supplementary material, table S1).

15The discrete changes in the size of icosahedral capsids are relatively larger for viruses with small genomes than for viruses with large genomes. We illustrate this principle using some DNA viruses whose T number is known and where the next highest T number can be predicted from other viruses that share the same capsomer fold (see electronic supplementary material, Supplementary Methods and table S2). These include eight tailed, icosahedral dsDNA bacteriophages with the HK97 fold and similar capsid molecular weight (around 40 kDa). To expand the range of T values, we included four single-stranded (ss) DNA viruses with T = 1 (no dsDNA viruses with this property are known). These viruses have a different fold, β-barrel fold, but their coat protein is of similar weight. The only transition that is hypothetical is from T = 16 (SPO1 and Syn9) to T = 19, i.e. no dsDNA phage with T = 19 is known, and we predicted this transition from the ‘3n + 1 rule’. For these calculations, the capsid has been approximated to a sphere and the relative volume increase has been calculated assuming the product of the radius and the reciprocal of the square root of T to be constant.

16We find that 75 per cent of the approximately 2000 known virus species have at least some gene overlap. The negative relationship between overlap proportion and genome length in RNA viruses (figure 2; linear regression r2 = 0.23, p = 0.006) reported previously also exists among DNA viruses (figure 3; linear regression, r2 = 0.38, p = 0.002). This relationship is found within all the constituent virus groups, e.g. ssDNA and dsDNA viruses (electronic supplementary material, figures S2 and S3), and within the two types of overlap: internal overlaps, where one gene is completely overlapped by a larger second, and terminal overlaps, where two genes overlap for part of their lengths—one upstream and one downstream (electronic supplementary material, figures S4 and S5).

17As summarized in figure 4 (electronic supplementary material, table S3), the mutation rates of RNA viruses are higher than those of DNA viruses. An important caveat here is that some DNA viruses have a longer genome than any RNA virus (compare figures 2 and 3) and are thus not comparable. There are estimates for only two DNA viruses of similar genome length to RNA viruses (Inoviridae and Microviridae). Nevertheless, these two values are lower than the mutation rate of any known RNA virus (we have found two lower mutation rate estimates for RNA viruses, in MLV and FLUBV, but these are both questionable because there are much higher estimates for the same or related viruses).

18Mutation rate appears to be a poor explanation of gene overlap in viruses because RNA viruses, despite their higher mutation rate, tend to have less genome overlap than DNA viruses of similar genome length—the opposite of the expectation if mutation causes overlap. The mean number of nucleotides in an overlap as a percentage of the genome length is 4.6 per cent for RNA viruses and 6.6 per cent for DNA viruses whose genome length is within the upper limit for RNA viruses (and 4.3% for DNA viruses if we include all genome lengths). Excluding families that lack overlap from these calculations does not change this result: the value is unchanged for small DNA viruses but increases to 6.1 per cent for RNA viruses and to 4.5 per cent for DNA viruses of all genome lengths. A lower value for overlap proportion in RNA viruses published previously by one of us was an error.

19As shown in figure 4, there is no obvious relationship between overlap proportion and mutation rate (linear regression, p = 0.50, n = 13). Introducing stepwise into an ANOVA (i) viral type (DNA or RNA) and (ii) genome length did not reveal significant interactions or a significant overall increase in the relationship between overlap and mutation (p > 0.26). The same outcome was obtained using—rather than a logarithmic—an arcsin transformation, which linearizes less effectively but allows inclusion of zero overlap values (n = 15).

20Separating virus families into ones with icosahedral capsids and those with flexible capsids (figures 2 and 3) shows that this negative relationship between overlap and genome length is actually derived only from the former group. Combining RNA and DNA viruses into one regression analysis, the relationship between overlap proportion and genome length is strong among families with icosahedral capsids (linear regression, r2 = 0.27, p = 0.001, n = 35) but there is no evidence for it among families with flexible capsids (linear regression, r2 = 0.02, p = 0.58, n = 19). We find the same result treating RNA and DNA viruses separately: in RNA viruses with icosahedral capsids r2 = 0.26 and p = 0.02, while in RNA viruses with flexible capsids r2 = 0.14 and p = 0.21; in DNA viruses with icosahedral capsids r2 = 0.81 and p < 0.001, while in DNA viruses with flexible capsids r2 = 0.08 and p = 0.60.

21This finding is not caused by the sample size differences. For both virus types, we sampled families with icosahedral capsids to give us the same number as families with flexible capsids and repeated the analysis 100 000 times: in only 9 per cent (RNA viruses) or 0.1 per cent (DNA viruses) of the replicates did the relationship decline to the same level as among families with flexible capsids, as measured by their p-value (excluding the outlying value for Plasmaviridae raises the DNA virus result only to 0.5%). As the data for RNA and DNA viruses are independent, these two probabilities (9% and 0.1%) can simply be multiplied.

22Most of the decrease in overlap proportion with increasing genome length is caused by overlaps becoming rarer rather than shorter. The evidence for this is that there is a strong negative relationship between the number of overlaps per unit length and genome length (linear regression, r2 = 0.37 and 0.38 in RNA and DNA viruses, respectively; electronic supplementary material, figure S6). By contrast, there is a much weaker negative relationship between mean overlap length and genome length (linear regression, r2 = 0.02 and 0.16 in RNA and DNA viruses, respectively; electronic supplementary material, figure S7).

23We are confident that this rarity of gene overlaps in larger viruses is not an artefact caused by larger viruses having longer genes, and hence fewer genes per unit length. This is because there is a very highly significant negative relationship between (i) the ratio of overlap number to gene number and (ii) genome length (linear regression, r2 = 0.21 and p = 0.007 in RNA viruses, r2 = 0.42 and p < 0.001 in DNA viruses; electronic supplementary material, figure S8).

24We find that the relationship between overlap proportion and genome length is strengthened by the removal of short overlaps (electronic supplementary material, figure S9; r2 values increase to 0.33 and 0.71 for RNA and DNA viruses, respectively). This is consistent with most large gene overlaps having evolved as a form of genome compression but with some short terminal overlaps having evolved to regulate gene expression (and others perhaps being neutral).

25Our findings are consistent with the inflexibility of icosahedral capsids constraining virus genome length, and gene overlap being a mechanism for acquiring new gene functions under this constraint.

26We do not find evidence for gene overlap having evolved as a response to the deleterious effects of mutation. However, we clearly need mutation rate estimates for more small DNA viruses, especially dsDNA viruses and those ssDNA viruses that are known to have a high rate of evolution (substitutions per site per year; ). Also, the expectation of more overlap among RNA compared with DNA viruses, owing to their higher mutation rates, could be masked by RNA viruses having a secondary structure and this further constraining overlap at synonymous sites.

27The hypothesis that gene overlap evolved in response to length constraint by the capsid assumes that there is a fitness cost to the virus in enlarging its capsid. We think this is a reasonable assumption because capsid enlargement can be expected to reduce the virus's reproductive output, which we can measure as the burst size or growth rate. Many studies have shown burst size/growth rate to be limited by the host cell's resources and producing a larger capsid will require more of these resources.

28We think it very unlikely that the inevitable misannotations in some RefSeq entries will have affected our conclusions. We recognize that detecting gene overlaps is difficult, with new overlaps recently being discovered using specially designed computer programs. However, the two critical points for our study are, first, that the trends we report are also found among small genomes (RNA viruses and the smaller DNA viruses) as well as among the larger DNA viruses, and it is among the latter where we expect most misannotations (because many overlaps will not have been confirmed experimentally); second, misannotation will be a source of random error that will obscure trends rather than create them, e.g. in our study the inclusion of provisional RefSeq genomes obscures completely the relationship between overlap proportion and genome length in dsDNA viruses (electronic supplementary material, figure S1); indeed, our study is a good example of the importance of data quality in comparative genomic studies. Similarly, uncertainty over the location of start codons might mean that short terminal overlaps are more likely to be misannotations than other overlaps; however, exclusion of these only strengthens the support for our findings (electronic supplementary material, figure S9), and we find the same relationship between overlap proportion and genome length when we consider only internal overlaps (electronic supplementary material, figures S4 and S5), which are generally unlikely to be misannotations because on average they are much larger.

29The relationship between overlap porportion and genome length is much weaker among icosahedral RNA compared with icosahedral DNA viruses (r2 values of 0.26 and 0.81, respectively). We suggest the following explanation. There is some evidence among RNA viruses that the genome length can physically influence the size of the capsid, perhaps because RNA virus genomes tend to be involved in capsid formation; by contrast, the genome of DNA viruses enters the already formed (pro)capsid. RNA viruses might thereby have a little more flexibility in expanding their genome length without overlap.

30This putative lower level of constraint among RNA viruses is consistent with some other observations. First, polyploidy (the packaging of more than one genome or genome segment copy) is known among RNA but not DNA viruses. Three of these examples are in families with flexible capsids (Orthomyxoviridae, Paramyxoviridae and Retroviridae) and one is in a family with icosahedral capsids, Birnaviridae. Second, most of the known examples of species being polymorphic for capsid size (so-called T number modulation) are among RNA viruses rather than DNA viruses.

31Another observation is consistent with capsid size influencing genome architecture. Many RNA virus species have segmented genomes (analogous to chromosomes), and a few of these package these genomic segments into separate capsids. It has been suggested that this phenomenon evolved in order to relieve packaging constraints and the need for overlap. For example, one can compare some members of the Comoviridae and Bromoviridae, which package genomic segments into separate icosahedral capsids and in which there is little or no overlap, with Tymoviridae, which have single icosahedral capsids and make extensive use of overlap. These are all plant viruses whose capsid size might be severely constrained by the need to pass through the plasmodesmata. Genome segmentation is rare among DNA viruses, but two of the three families in which it occurs, Nanovirus and Geminiviridae, are also icosahedral plant viruses in which at least some members have genomic segments that are packaged separately (the third family is the non-icosahedral Polydnaviridae).

32There are of course other factors involved in the evolution of gene overlaps in viruses, e.g. there are overlaps in non-icosahedral viruses and in other organisms: (i) selection for faster replication could lead to gene overlap: if the small genome of some viruses is the result of selection for faster replication (small genomes being quicker to copy), then we might expect small viruses to have more overlap because overlap allows more proteins to be coded for by a given genome length. There is a more general evolutionary model to explain genome compression which incorporates mutation rate, population size and replication rate, but our comparative data cannot test this model. Selection for replication speed has also been proposed as an explanation for the high mutation rate of RNA viruses, via a possible trade-off between copying speed and copying fidelity, and this area deserves more attention, (ii) proteins and protein regions created de novo by overlapping may have novel chemical and physical properties. This is suggested both by contemplating the effects of a resulting shift in codon usage bias and by observing that such genes tend to have unusual protein structure and composition. Nevertheless, we find it difficult to explain the striking relationship between overlap and genome length if overlaps evolved primarily in order to create proteins with novel chemical or physical properties—why should this be more important for small viruses than for large viruses? Also, we do not know the relative importance of gene overlap in creating the genomes of today's RNA and DNA viruses. Large dsDNA viruses have acquired many genes via gene duplication and horizontal transfer from the host, while there are very few examples of these processes known among RNA viruses; however, the comparison needs to be between RNA viruses and DNA of comparable genome length, and this has yet to be done, and (iii) some overlaps may be selectively neutral. It is likely that overlaps evolved from the translation of novel or extended open reading frames (ORFs) created by the mutational gain or loss of start and stop codons, respectively, although the actual molecular mechanisms involved are varied. Subsequent mutations could lengthen these ORFs and new functions could be acquired gradually by the novel polypeptides. Thus, initial short overlaps may be essentially neutral. We need to test this model by reconstructing the appearance and changes in length of overlaps on viral phylogenies, which could be combined with examining changes in capsid structure, e.g. we predict that short overlaps precede longer ones and increases in capsid size follow gene acquisition.

33There has been a search for universal factors that influence an organism's genome length, and—beyond a commonly observed genome reduction in symbionts—there is little consensus at the moment. We believe that our analysis of one possible form of genome compression points to a taxon-specific factor, namely the capsid, and casts doubt on the role of a more general phenomenon, namely mutation. We also believe that our study is a good example of contingency in evolution: natural selection acted on capsids favouring the icosahedral structure because of its stability and economical design, the consequence of which was a fixed volume of the interior. This led to the proliferation of overlaps not necessarily as the best possible solution to increase the genome coding capacity, but as the only one possible. Where this constraint is less stringent, less overlap is present. In effect, the capsid poses an engineering problem for the creation of genomic novelty, and gene overlap is the way around it.