However, Grantham, although suspecting it was nucleic acid structure that was conserved at synonymous sites, focussed only on mRNA structure. Indeed, decades later, while affirming that "the underlying biological function disrupted by these [synonymous] mutations is unknown, but it is not related to the forces generally believed to be -- shaping the evolution of synonymous sites,"  Lawrie et al. [1] still held that "a strong possibility remains that the function underlying the strong constraint at synonymous sites is related to mRNA structure." Likewise, in 2013 Lind and Andersson [11] concluded that "the deleterious effects of synonymous mutations are not generally due to codon usage effects, but that mRNA secondary structure, is a major fitness constraint," and Park et al. [12] claimed "a major role of natural selection at the mRNA level in constraining protein evolution."

Of course, structure conservation at the mRNA level means that conservation is also at the DNA level of the corresponding genes. Which is primary and which is secondary? Since, the potential of duplex DNA to adopt stem-loop configurations is often better developed in non-genic regions and introns, than in exons [13-16], the pressure for conservation could have arisen primarily at the genome level (be the genome itself DNA or RNA) . Indeed, in 1986 Bernardi and Bernardi [8] referred to selection acting on the "genome phenotype" rather than on the conventional phenotype. This would be consistent with a selective pressure relating to recombination [6]. Mutations likely to affect recombination are of two general forms, those affecting the proteins that mediate recombination (conventional phenotype), and those affecting their targets - DNA itself (genome phenotype). The former mutations would be localized to genes (mainly non-synonymous mutations), the latter would be genome-wide (synonymous and non-synonymous when within protein-encoding regions), and could involve DNA structure.

The forces manifest when such mutations occur, should also operate in organisms with RNA genomes. Because of high mutation rates and high clinical interest, studies of retroviral evolution now provide well-documented exemplars of the general forces operating on genome sequences [17]. These forces can either promote or decrease recombination. In the latter case, new species can arise. While mechanisms of speciation are controversial, there is general agreement that "mutation is crucial in speciation because reproductive barriers cannot be generated without mutations" [18]. But, for RNA viruses in general, Simon-Loriere and Holmes [19] have argued that "there is little evidence that recombination is favoured by natural selection to create advantageous genotypes or purge deleterious mutations, as predicted if recombination functions as a form of sexual reproduction." However, as reviewed here, at least in the case of HIV-1, studies of RNA structure and base composition suggest otherwise.

2. Recombination, rejuvenation and conservation of RNA structure

At its most fundamental level, sex is recombination [20, 21]. The conjugation required for the meeting of paternal and maternal human gametes is an elaborate preliminary to the final meiotic meeting of parental genomes in the gonads of their offspring, where recombination occurs. The early microscopists who witnessed this final meeting described it as a "conjugation of the chromosomes" that was necessary for a rejuvenating "interchange of substances"

"The conjugation of the chromosomes in the synapsis stage may be considered the final step in the process of conjugation of the [parental] germ cells. It is a process that effects the rejuvenation of the chromosomes; such rejuvenation could not be produced unless chromosomes of different parentage joined together, and there would be no apparent reason for chromosomes of like parentage to unite."

In contrast to humans, retroviral sex is quite elementary [23]. Yet, like humans, retroviruses are diploid. This diploidy in HIV-1 can be heterozygous due to the viral strategy of mutation to near oblivion, so countering host defences. The degree of heterozygosity in two HIV genomes that are copackaged within an infectious retrovirus particle, if below the sequence difference threshold above which recombination is inhibited (see later), will allow viral rescue by recombination. Thus, within the next host cell, the two partially crippled genomes can repeatedly recombine to generate a rejuvenated form that will successfully colonialize the vulnerable population of host cells (usually T4 lymphocytes) [24]. This implies that the virus will have accepted mutations supporting the ability to recombine. Indeed, if RNA structure were required for recombination, then HIV-1 genomes might display higher conservation of RNA structure than of the encoded proteins. Evidence for this postulate was presented in 1995 [25] and is now well supported [26-30]. Furthermore, beyond its recognized regulatory roles, viral RNA structure can be related to recombination and speciation [15, 16, 25].

Apart from this, HIV-1 genomes have high and stable levels of the purine A [31]. This A-richness has been related to the phenomenon of purine-loading and has implications for RNA structure [32, 33]. In keeping with current conventions [34], "high A%" here implies a high frequency of the base A. "High AT%" implies a high frequency of the combination of bases A and T, without specification of which of the two bases makes the major contribution. The same considerations apply to high values for GC% and AG% ("purine loading").

3. Purine-loading

The A-bias in HIV-1 is a characteristic of coding regions, but not of the non-coding long terminal repeats (LTRs). This parallels the selective "purine-loading" (AG% enrichment) of the encoding (mRNA synonymous) strands of the exons of genes of most biological species. Since purines do not pair with purines, this would militate against host self RNA-self RNA interactions. A purine-loaded virus would be less likely to form base-paired RNA duplexes with the host RNAs that appear when the genomes of stressed cells are pervasively transcribed (i.e. this would prevent self RNA-not-self RNA interactions) [34]. Preventing formation of double-stranded RNA (dsRNA) should be advantageous to the virus, since segments of dsRNA can provide alarm signals to the host, so alerting immunological defences [35].

Purine-loading also provides a rationale for interruption of proteins by low complexity segments corresponding to purine-rich codons [36, 37]. This implies function at the nucleic acid level. In other words, removing such segments from proteins should not affect their function. Thus, in 2011 Muralidharan et al. [38], from studies of a P. falciparum protein with an asparagine repeat sequence (corresponding to purine-rich codons GAY), concluded that "the asparagine repeat is dispensible for protein expression, stability and function." A similar observation has been made for a repeat within the purine-rich mRNA encoding the EBNA-1 protein, which normally maintains the Epstein-Barr virus (EBV) in a deeply latent state with its many pyrimidine-rich mRNAs unexpressed. In this way EBV should evade host immune defences [35, 39, 40].

Irrespective of whether avoidance of dsRNA is the correct adaptive explanation, purine loading is a fact. A-loading by HIV-1 could be an extreme expression of this [35]. Purine-loading can influence amino acid composition. Since the extra purines generally locate to third codon positions (the main site for synonymous mutations), it is likely that purine-loading has driven amino acid composition, rather than the converse [41]. The purines would be expected to occupy, and are indeed found in, the loop regions of the highly structured HIV genome [26, 33]. Indeed, in 2013 Mayrose et al. [42] noted pervasive synonymous site conservation in HIV-1, and Zanini and Neher [43] found that most synonymous mutations disrupt base pairing in HIV RNA stems and suggested direct fitness effects of the stem-loop structures.

Values for AG% and GC% are reciprocally related, and purine-loading is achieved mainly by replacing C with A, leaving T and G relatively constant (so avoiding runs of Ts and Gs that can be detrimental). Thus, there can be conflict between 'AG pressure' and 'GC-pressure' [34, 41]. To appreciate why HIV-1 is A-loaded, not G-loaded, we return to the role of GC%.

4. Two species in one cell differ greatly in GC%

Consider three examples of pairs of viral species, where members of each pair might occupy the same host cell. (1) An early report on species differences in GC% [44] noted that, while there was no clear relationship between the GC% values of virus and host, two insect viruses (polyhedral and capsule) with a common larval host (spruce budworm) differed dramatically in their GC% values (38% versus 51%). In this example it was uncertain whether these viruses could have shared a common cell within their larval host.

However, in 1991 Schachtel et al. [45] noted for humans, that two neurotropic alphaherpesviruses also differ dramatically in GC% (46% versus 68%). The differences were dispersed, being present in both non-coding and coding regions and, in the latter case, mainly affected the third bases of codons (consistent with the close similarity in amino acid sequences of the viral proteins). But, as in the case of purine-loading (see above), the differences were sometimes sufficient to change amino acids (e.g. arginine to lysine). It was proposed that the differences were adaptive in that there would be less competition for host resources:

"More specifically, it is proposed that HSV1 and VSV avoid competition for host resources such as nucleic acid precursors and aminoacyl-tRNAs through divergent base composition and codon usage. This may enable these two closely related viruses to coexist in the same host species and even to multiply simultaneously in the same cells."

In 1994 a similar explanation was advanced by Bronson and Anderson [46] who contrasted the AU-richness of HIV-1 (42% GC), with the GC-richness of HTLV (53% GC). They supposed that the intracellular environment would consist of a variety of metabolic "ecological niches." Two viruses that coinfected the same host cell would avoid competition with each other by making different metabolic demands on NTP pools. There is, however, another interpretation.

5. Coinfectants either blend or speciate

The wide GC% differences between coinfecting viruses can be seen as part of the speciation process [15, 16, 25]. Two species of virus derived from a common ancestral species, which coexist synchronously within the same host cell (sympatry), have the opportunity to recombine. Due to this 'blending inheritance,' their reproductive isolation would then be lost and they would mutually destroy each other as independent species. To the extent that retaining their differentiation as independent species is advantageous, there would have been a strong selective pressure on their nucleic acids (i.e. on their genome phenotypes) to evolve to prevent recombination. If differences in GC% could achieve this (see later), then they would have been mutually driven to different GC% poles. If HIV-1 were the first to A-load (i.e. accept mutations to A), then it would have been driven to the low GC% pole. HTLV would have been driven to the high GC% pole. Since, to arrive at high GC% values, organisms preferentially replace A with C [41], HTLV would have foregone the advantages of purine-loading and adopted an alternative evolutionary strategy [35]. With its preference for A, HIV-1 recombinationally drove HTLV-1 away.

We should note that, like the GC-rich EBV (see above), HTLV-1 is deeply latent and many carriers are symptom-free. Also like EBV, the duplex HTLV-1 provirus avoids transcribing its many pyrimidine-rich RNAs (controlled by the 5' LTR region). This is achieved by expressing a protein (HBZ) with a role analogous to that of the EBNA-1 protein (the latter encoded by a purine-rich mRNA). HBZ mRNA is transcribed from the antisense DNA strand as a purine-rich RNA species (controlled by the 3' LTR) [47].

6. FORS-D analysis and SHAPE illuminate RNA structure

New technologies have greatly assisted the understanding of HIV RNA structure. In 1988 Le et al. [48, 49] used novel energy-minimization folding algorithms to show that RNA regions involved in intra-molecular base pairing tend to be more evolutionary stable. This stability of nucleic acid folding into stem-loop conformations is expected to be influenced by base composition, being low in low GC% organisms such as HIV-1 (since there is less opportunity for base-stacking with pairing between Gs and Cs). However, GC% values tend to characterize whole genomes or large genome sectors (see above). Base order is a character that critically affects local structure. A method of dissecting out the base order-dependent component of the folding energy - "folding of randomized sequence difference" (FORS-D) analysis (for details see [50]) - revealed many evolutionarily conserved structural elements in the HIV RNA genome; these were separated by less structured variable regions likely to be under positive Darwinian selection [25, 29, 42, 43].

These observations were recently confirmed and extended by "selective 2'-hydroxyl acylation analysis by primer extension" (SHAPE) [26]. While certain conserved structures corresponded to established regulatory elements, others corresponded to the peptide linkers between the protein domains of polyproteins. The results were interpreted in terms of "another level of genetic code" so that "higher ordered RNA structure directly encodes protein structure." Recombination was not mentioned. Others agreed, stating that "this novel component of the genetic code represents the strongest determinant of conservation" [29]. Again, there was no suggestion of a relationship to recombination.

7. Recombination depends on structure and species

If the degree of heterozygosity of two HIV-1 genomes, copackaged within an infectious retrovirus particle, is insufficient to generate sequence differences of a magnitude that inhibits recombination and fosters speciation (see below), then, by virtue of similarities in sequence and structure, the two genomes can generate recombinants with properties more advantageous than those of the originating parental genomes. But HIV-1 itself, by virtue of its high mutation rate, explores sequence space in years, to an extent that would take its host millions of years. Thus, as sequence differences increase, a differentiation of HIV-1 into "quasispecies" (up to approx. 15% intrasubtype variation), "subtypes" (approx. 15% - 30% intersubtype variation) and "groups" (>30% intergroup variation), is recognized. How efficiently can recombination have checked this variation? Can differences in recombination between members of these various categories, although mechanistically copy-choice (template switching without strand breakage), guide our understanding of speciation in more complex organisms where recombination is likely to involve DNA strand breakage [15, 16]?

Recombination in HIV-1 begins with the formation of RNA dimers by means of complementary loop-loop "kissing" interactions between "dimer initiation sequences" (DIS) that are part of stem-loop secondary structures in the Ψ (packaging signal) region of the genome. A similar DNA "kissing" may be involved in recombination in higher organisms [15, 51]. A lack of base complementarity between HIV-1 DIS loops decreases recombination, but does not eliminate it. This suggests that other parts of the genome can assist dimer formation. Upon transfer to a new host cell, copy choice recombination occurs between the colocalized genomes in HIV-1 homodimers or heterodimers [52, 53].

Since a species is defined by its recombinational (reproductive) isolation relative to other species [16, 18], it is important to note that, albeit rare, recombination can occur between members of different HIV-1 groups [54]. This defines them as belonging to the same species. On the other hand, although HIV-1 and HIV-2 can coinfect an individual, no recombinants have been detected, so they are different species. HIV-1 intrasubtype recombination is more frequent than intersubtype (intragroup) recombination, which is more frequent than intergroup recombination [55]. Thus, after introduction of synonymous substitutions, it was found that 5%, 9%, and 18% sequence differences corresponded, respectively, to 35%, 74% and 95% decreases in recombination frequency [52]. Others, with a different assay, found 9% and 18% sequence differences corresponded to 67% and 90% decreases in recombination [53]. These values are consistent with previous FORS-D studies (see below), which postulated that HIV-1 secondary structure is conserved because, in addition to regulatory roles, it plays a critical role in recombination [25]. Consistent with this, in 2011 Simon-Loriere et al. noted [28]:

"Strong disparities [in recombination] were observed for comparable degrees of sequence identity in conserved regions, indicating that -- parameters other than the level of sequence identity modulate -- recombination. -- [R]egions of the genomic RNA with a high proportion of residues involved in the formation of secondary structure contained significantly more [recombinational] breakpoints. The extent of RNA structure along the HIV genome seems to provide us with a relatively accurate picture of the pattern of recombinant genomes generated by the mechanism of recombination."

Acknowledgements