The evolution of X chromosome inactivation in mammals: the demise of Ohno’s hypothesis?

Abstract

Ohno’s hypothesis states that dosage compensation in mammals evolved in two steps: a twofold hyperactivation of the X chromosome in both sexes to compensate for gene losses on the Y chromosome, and silencing of one X (X-chromosome inactivation, XCI) in females to restore optimal dosage. Recent tests of this hypothesis have returned contradictory results. In this review, we explain this ongoing controversy and argue that a novel view on dosage compensation evolution in mammals is starting to emerge. Ohno’s hypothesis may be true for a few, dosage-sensitive genes only. If so few genes are compensated, then why has XCI evolved as a chromosome-wide mechanism? This and several other questions raised by the new data in mammals are discussed, and future research directions are proposed.

Introduction

Since its origin ~180 MY ago [1, 2], the human Y chromosome has lost ~97 % of the genes originally present on that chromosome [3] (Fig. 1a). This massive gene loss on the Y resulted in an imbalance of X-linked gene dose in males with their single X chromosome compared to females with two Xs (Fig. 1b). Such dosage imbalance was probably deleterious as in humans, dosage imbalance caused by chromosomal aneuploidies (e.g., monosomy, trisomy) of autosomes the size of the X chromosome are lethal. Specific mechanisms have evolved in animals to compensate for this gene dose problem in males, and the solution found in mammals appears to be a complicated one. After Susumu Ohno and Mary Lyon discovered independently female X-chromosome inactivation (XCI) in mammals [4, 5], Ohno proposed that dosage compensation in mammals evolved as a two-step mechanism with (1) a twofold expression increase of the X chromosome in both sexes, which solves the gene dose imbalance problem in males, and (2) inactivation of one of the two X chromosomes by XCI in females to restore optimal dosage [6] (Fig. 1b).

Fig. 1.

Sex chromosome and dosage compensation evolution in mammals. a Sex chromosome evolution in mammals. Sry, the male-determining gene, initiated the evolution of the sex chromosomes from a pair of autosomes. The proto-X and proto-Y stopped recombining at a region including Sry and probably other genes, thus forming X- and Y-specific regions. These regions grew larger during evolution through additional recombination suppression events (probably inversions on the Y chromosomes) and gradually diverged. Pseudo-autosomal regions (PARs) are remnant of the autosomal ancestry of the sex chromosomes. In its non-recombining male-specific region, the Y has lost most of its genes because of degenerative processes collectively known as Hill-Robertson effects [57, 73]. b The evolution of dosage compensation in mammals, as hypothesized by Ohno [6], formalized later by Charlesworth [74, 75], and explicitly modeled by [48]. The gene loss on the Y implies dosage imbalance between the sex chromosomes and the autosomes in males. In the first step of Ohno’s hypothesis, expression of the X chromosome is doubled in both sexes; proper dosage is restored in males, but is now twice the dosage of autosomes in females. In the second step, the inactivation of one of the two Xs in females evolves in order to get the dosage of the X back to autosomal level. The predicted values of both the expression ratios X:AA and X:XX (see Box 1 and text for more details) are shown at the bottom of panel b. Expression on the autosomes may have changed during evolution, hence the less precise prediction for the X:AA ratio than for X:XX one (Box 1), as emphasized by the tilde symbol

XCI has been widely studied both at the mechanistic and evolutionary level (reviewed in e.g., [7–15]). In placental mammals, the silencing of the inactivated X (Xi) is established through epigenetic signals involving a long non-coding RNA called Xist and a number of cis and trans factors affecting its transcription. In particular, this involves Rnf12 (main activator of Xist), and Rex1 (main inhibitor of Xist) [16]. RNF12 is an E3 ubiquitin ligase that binds REX1, which triggers its degradation by the ubiquitin pathway. This results in Xist being activated or not depending on RNF12 dose, which differs between male and female as Rnf12 is an X-linked gene [16]. XCI is initialized in a particular region of the X chromosome (called the X inactivation center, XIC) where the Xist gene is located. It then spreads across the X with Xist being transferred directly from XIC to distal sites across the X chromosome that are defined not by specific sequences but by their spatial proximity in the nucleus to XIC [17, 18]. XCI is established early during development but there are substantial mechanistic and timing differences among species [13, 14]. XCI has evolved region by region on the X chromosome, starting with the region where recombination between X and Y ceased and Y degeneration started first, and encompassing more and more of the X chromosome as recombination suppression progressed [19]. XCI is considered a chromosome-wide phenomenon, but interestingly, 10–15 % of the genes on the human X chromosome escape XCI. This includes not only the genes in the still-recombining portion of the sex chromosomes (“pseudoautosomal” regions, PARs), but also blocks of genes in the X-specific region including X-linked genes with a still active Y homolog [11, 19, 20]. XCI has probably evolved early in the evolution of the mammalian sex chromosomes, even though Xist has emerged only in the placental lineage from a protein-coding gene called Lnx3 [21].

For a long time, data on the two-fold expression increase on the X chromosome was lacking. Thus, the first step in Ohno’s hypothesis has remained very speculative, but at the same time was widely accepted in the scientific community. Only when chromosome-wide analysis of gene expression became possible has Ohno’s hypothesis started to be tested. The first studies using microarray seemed to support Ohno’s idea of X expression doubling in both sexes [22–26]. However, the first study using RNAseq, a next-generation sequencing technology (NGS) to study gene expression, found no evidence for Ohno’s hypothesized first step [27]. An avalanche of papers followed, some supporting Ohno and others contradicting him [28–35]. The aim of this review is to explain this ongoing controversy, to show that despite the controversy, a novel view on dosage compensation evolution in mammals is starting to emerge and to highlight the fundamental questions that remain to be answered.

The controversy about testing Ohno’s hypothesis

In mammals, an early study found some support for Ohno’s hypothesized X expression doubling by showing that an autosomal gene in laboratory mouse strains exhibited a doubling of its expression following translocation to the X chromosome in Mus spretus [36]. However, it was only later with the advance of the microarray technology that Ohno’s hypothesis could be tested with many genes (Table 1). This was done by comparing the global expression on the X chromosome to that on the autosomes and computing the X:AA expression ratio (Box 1). In accord with Ohno’s hypothesis, a mean X:AA expression ratio close to one in male as well as female tissues from several mammalian species including human, macaque, mouse, and rat was obtained [22, 23]. Moreover, similar X expression in both sexes was found in human and mouse. Several other microarray studies were conducted and they all supported Ohno’s hypothesis. However, the accuracy of the expression level estimates from microarray data was later criticized (Box 1).

Table 1.

Summary of recent studies testing for Ohno’s hypothesis

Human X:AA	Mouse X:AA	Evidence for global hypertranscription	Expression data	Dataset filtering	References
0.9	1	Yes	Microarray	d	[22]
–	1	Yes	Microarray	d	[23]
0.5	0.2	No	RNA-seq	Excluding same proportion of lowly and highly expressed genes from A and Xb	[27]
0.9 (0.6)	0.8 (0.5)	No if procedure for data filtering is changed	Microarray	(same proportion of genes from A and X) d	[68]
0.9	0.9	Yes	RNA-seq and PolII occupancy	d	[28]
0.5	–	No	RNA-seq	Grouped by expression levelsd	[29]
0.9	0.8	Yes	RNA-seq	d	[30]
–	1	Yes	Microarray	c	[31]
–	0.8	Yes	RNA-seq and PolII occupancy	c	[35]
0.7 (0.9)	–	No, except for genes involved in complexes ≥7 proteins	RNA-seq	(genes involved in complexes ≥7 proteins) c	[32]
0.5/0.5a	0.4/0.5a	No, evidence for hypotranscription for some interacting genes	RNA-seq	Conserved in all amniotesc	[33]
0.5/0.5a (0.9) a	0.3/0.4a	No, except for genes involved in complexes ≥7 proteins	RNA-seq	(genes involved in complexes ≥7 proteins)b	[34]

() values given in brackets and in bold correspond to the data subset indicated in the dataset filtering column (also in bold)

^aX:XX ratio

^bAll genes

^cExpressed genes (FPKM > 0)

^dActively expressed genes (FPKM ≥ 1 or FPKM ≥ 3 or detected in ≥95 % of microarray samples)

In 2010, the first study using RNAseq, a NGS technology supposed to give much better estimates (Box 1), reported a median X:AA ratio close to 0.5 in a variety of human tissues (both from male and female) and an even lower ratio in mice, challenging Ohno’s hypothesis [27]. A year later, several other studies using more RNAseq data were published (Table 1). The 2010 study was criticized for having included genes with no or very low (probably noisy) expression in their analysis. The X chromosome includes more tissue-specific (mostly testis-specific) genes than the autosomes. This means that for instance in liver tissue, many testis-specific genes on the X have no expression and because the X has more of these genes than the autosomes, the X:AA ratio is reduced. When only the genes expressed in one tissue were included to compute the X:AA ratio of that tissue, X:AA ratios were much closer than 1 in both humans and mice [28]. Moreover, ChIP-chip data in mice showed a relatively higher occupancy of an active form of RNA polymerase II (PolII) for highly expressed X-linked genes compared to highly expressed autosomal genes, giving more support to the idea of X hyperexpression proposed by Ohno [28]. Quite strikingly, the same issue of Nature Genetics included a total of five articles reporting tests of Ohnos’ hypothesis, some reporting a X:AA close to 1 [28, 30], another re-analyzing microarray data and confirming a X:AA of 1 [31]. A paper published in another journal also reported higher PolII occupancy and more active histone marks on the active X chromosome than on the autosomes [35]. Ohno’s hypothesis seemed to regain support.

However, in their reply to these papers, the 2010 paper authors criticized excluding genes with no expression as being arbitrary [29]. They noticed that testing Ohno’s hypothesis using X:AA ratios will work only if global expression of the proto-X and the autosomes was initially the same—an assumption that was never tested. As global expression varies even among autosomes, it is possible that the expression of the proto-X differed from that of the other chromosomes even before Y gene loss had started. If this were the case, the present-day X:AA is not expected to be one even if Ohno’s hypothesis is correct and expression doubling took place during evolution. A much better way of testing Ohno’s hypothesis is to compare the present-day expression of the X chromosome to its ancestral expression, the X:XX ratio (Box 1). This has recently been done by focusing on the genes that have 1:1 orthologs between chicken and humans using an amniote-wide RNAseq dataset [33, 37]. Importantly, these ‘old’ genes were initially present on the proto-X chromosome and are those that should show patterns of dosage compensation. By contrast, such patterns are not expected for the many “young” genes gained late in the evolution of the X chromosome [38]. The ancestral expression of the “old” genes was estimated using the expression of their autosomal orthologs in chicken, and the X:XX was found to be 0.5 for placental species including human, chimpanzee, bonobo, gorilla, orangutan, macaque, and mouse. In marsupials (opossum), however, the X:XX expression ratio was found to be one. Interestingly, this study also showed that the assumption that expression was similar in the proto-X and the autosomes underlying all the studies using X:AA was actually correct (XX:AA ≈ 1 and AA:AA ≈ 1). Even measuring the X:AA ratio for the “old” genes returned a value of 0.5 as noted previously [29]. Including or excluding the tissue-specific genes did not change anything in this pattern (as tissue-specific genes are mostly “young” genes). This was later confirmed by an independent analysis of the same dataset, which found a X:XX of 0.5 for all placentals and also marsupials tested [34]. For unknown reasons, the results for marsupials differ in both studies. Importantly, X:AA ratios of 0.5 were confirmed using protein abundance, which suggests that the observed patterns are robust to the method of measuring expression [27, 34].

Dosage compensation of a minority of dosage-sensitive genes

The latest tests of Ohno’s hypothesis using ancestral X expression seem thus to reject it (Fig. 2a), and even the tests using X:AA expression ratio do not fully agree with Ohno’s hypothesis (in [28, 30, 34] X:AA ratios are lower than one when including poorly to moderately as well as highly expressed genes in the analysis, as noted in [32]). Are the sex chromosomes left with dosage problems in mammals? In humans, removing one chromosome is usually lethal. Y degeneration however took millions of years and was gradual, and the dosage problems may not be as severe as in an instantaneous loss of a chromosome. Also, buffering mechanisms that can partially compensate for the loss of a chromosome do exist [39]. Another possibility is that dosage compensation in mammals is a “half-full, half-empty glass” problem, with some genes being compensated but not all [40]. Looking at all the genes at the same time returns a X:AA ratio between 0.5 and 1, which some consider consistent with Ohno’s hypothesis (half-full glass) and others inconsistent with the same hypothesis (half-empty glass).

Fig. 2.

Dosage compensation of a minority of human X-linked genes. a X:AA and X:XX ratios for humans are shown (find more details about these ratios in the text and in Box 1). Results are shown for (1) all human genes or including only genes with a minimum expression level of FPKM > 1 (All genes), (2) 1:1 orthologs between human and chicken, for which ancestral expression could be computed using expression data in chicken (Box 1), considering all of them or only those with a minimum expression level of FPKM > 1 (1:1 orthologs with chicken), and (3) genes involved in large protein complexes (with 7 or more proteins) that are likely dosage-sensitive (Dosage-sensitive genes). Boxplots of the X:AA or X:XX medians for different tissues are shown. Extreme outliers can be seen for “Expressed genes” (“All genes” category), they correspond to Brain (highest ratio) and lung (lowest ratio). Data for preparing the “All genes” part and the X:AA of “Dosage-sensitive genes” part are from [32] and are based on 12 tissues. All the other boxplots were obtained from ten tissues in [34]. The blue dashed lines indicate the expected ratios with global dosage compensation (1), and without any dosage compensation (0.5), see text and Fig. 1b for more details. b Sketch summing up the differences in dosage compensation status and mechanisms among the genes on the X chromosome. Most of the genes on the Xi are inactivated, except for the PARs and some XCI-escapees, and XCI is thus a global process. The hyperexpression, on the contrary, appears to be a local process affecting only the dosage-sensitive genes. Dosage compensation through hyperexpression and XCI as envisioned by Ohno thus only affects dosage-sensitive genes [32]. Some dosage-sensitive genes are compensated through another mechanism, namely down-regulation of their autosomal partners, as shown for some genes involved in protein–protein interaction networks [33]

Many genes are known to be haplosufficient or dosage-insensitive [41], i.e., it is not lethal to lose one functional copy of those genes. For instance, it has been recently shown using a theoretical approach that X-linked genes involved in metabolic networks can easily lose a copy without a significant effect on the flux of the network and on fitness, especially for networks with many steps [42]. Dosage compensation should evolve for haploinsufficient or dosage-sensitive genes only [40, 43]. No exhaustive list of mammalian dosage-sensitive genes is available, but there are some known good candidates. The genes encoding proteins involved in complexes (protein-complex genes) are among those candidates. Stoichiometry of the components of a complex is required for its proper folding and functioning [41]. Using human RNAseq data, two of us studied the X:AA expression ratio within complexes, and found that the X:AA for large complexes (seven or more proteins) is around one [32]. Also, it was shown that X expression increase (and not autosomal expression decrease) explains this result by comparing X and autosomal expression among large and small complexes, and finding similar autosomal expression and increased X expression in large versus small complexes [32]. This was later confirmed by using ancestral expression levels and looking at each complex’s X:XX and AA:AA ratio [34]. Other dosage-sensitive candidates are genes involved in protein–protein interaction networks. Julien et al. [33] studied those genes and found that in some cases, autosomal genes have evolved reduced expression following their X-linked partners, another way of achieving dosage compensation. Dosage-sensitive genes on the X chromosome are thus dosage-compensated, some showing the hyperexpression proposed by Ohno [44] (Fig. 2; Table 1). However, these genes represent a minority of the X-linked genes (even though there are probably some unidentified ones on the X [45]), which explains the “half-full, half-empty glass” problem when studying Ohno’s hypothesis with all X-linked genes.

The mechanisms ensuring dosage-sensitive genes to match the expression of their autosomal partners in complexes or in protein networks are unknown. In some cases, expression of the X-linked genes was increased [32, 34], in other cases expression of the autosomal genes was decreased [33], which suggests that these mechanisms evolved on a gene-by-gene basis. Using ChIP-chip or ChIP-seq in mice, a higher RNA Polymerase II occupancy was found on the X chromosome than on the autosomes [28, 35]. More epigenetic marks characteristic of actively transcribed genes were found on the X chromosome compared to autosomes [35]. Interestingly, these trends were only found for highly expressed genes (also noted in [40]), of which dosage-sensitive genes may represent a substantial fraction [32, 46, 47]. However, it would be important to explicitly compare epigenetic patterns of dosage-sensitive versus other genes in order to gain a better understanding of the different mechanisms of adjusting expression of dosage-sensitive genes.

The origin of XCI and the early steps of dosage compensation evolution

Only a minority of genes show clear evidence of dosage compensation in mammals. If XCI initially evolved to counteract hyperactivation of the X chromosome in females as proposed by Ohno, it is not clear why XCI is global and affects the majority of the X-linked genes when hyperactivation is local and affects only a few genes. Of course, XCI may be global for unknown mechanistic reasons and its effect on many haplosufficient/dosage-insensitive X-linked genes may simply be neutral. It is also possible that XCI evolved for a completely different reason in the first place, and was only later in evolution recruited for dosage compensation (Fig. 3). This idea is somewhat supported by a population genetic model that has shown that XCI can evolve under Ohno’s hypothesis under quite restricted conditions only [48].

Fig. 3.

Steps in the evolution of XCI. Major events the evolution of XCI in placentals and marsupials are shown in the tree of amniotes. XCI, probably pXCI (shown in green), evolved early in the evolution of XY chromosomes, as shown here. However, independent XCI evolution in placental and marsupial lineages cannot be ruled out (see text). Also shown is how the mechanism of XCI was later refined independently in placentals with the evolution of the lncRNA Xist (from the protein-coding gene Lnx3) and Rex1 and became rXCI (shown in brown), and in marsupials with the evolution of the lncRNA with Xist-like properties Rsx (which has not evolved from Lnx3). In some placentals, pXCI is found as well (in early embryo and extra-embryonic tissues); it is not known whether pXCI has re-evolved or has been conserved, hence the question marks. Two major players in the evolution of XCI, Lnx3 (parent of Xist) and Rnf12, were probably close to Sox3 (parent of Sry) in the proto-X, and the same small region has apparently been involved in both the evolution of sex determination and XCI (see text). Birds and monotremes sex chromosome systems evolved independently from that in therians, and serve as outgroups here

Such an alternative theory for the evolution of XCI was proposed by Haig [49, 50]. According to his “parental antagonism model”, XCI initially evolved not as a means of dosage compensation, but as a silencing mechanism of growth-inhibiting genes on the X chromosome during embryonic growth (see Box 2 for details). The premise of this hypothesis is the theoretical expectation that the X chromosome is enriched for growth inhibiting genes [51]. Whether or not this premise holds is still unknown, but several lines of evidence are consistent with this idea. It has been shown that the number of X chromosomes affect the speed of early development before XCI is established, with X0 and XY mouse embryos developing faster than XX embryos (reviewed in [15]). In marsupials, H19X, a long coding RNA involved in regulating placenta growth has been found next to Rsx, the long coding RNA mediating XCI, which suggests a possible link between imprinting, placenta and XCI [52]. More work is needed but it is possible that the imprinted region was initially large enough for a mechanism such as XCI affecting several neighbor genes at once to evolve, paving the way for global XCI as we know it today. Another attractive feature of Haig’s hypothesis is that it predicts imprinted paternal X inactivation (pXCI) rather than random X inactivation (rXCI) to be the primary form of XCI. pXCI is found in marsupials, and also in extra-embryonic tissues and the early developing embryo of some placentals [7, 13]. Overall, the observations fit with the idea that pXCI was ancestral to rXCI (see next section for more details). However, parallel evolution of pXCI in marsupials and some placentals cannot be ruled out at this point. The parental antagonism model also predicts that XCI should have evolved in groups where parental conflicts about maternal resource allocation to embryonic growth are strong. This is indeed the case in placental and marsupial mammals. Recent data suggest that chromosome inactivation might also affect the bird Z chromosome and the monotreme X chromosomes where parental conflicts may be weaker [53], but this observation needs to be confirmed before taken as evidence against the parental antagonism model.

Another possibility is that the origin of XCI is connected to sex determination [12]. Sox3, the gene that gave rise to Sry, induces the development of testis when overexpressed, which suggests that evolving appropriate dosage of Sox3 may have been a crucial point in establishing sex determination through the Sox3/Sry gene pair in mammals [12]. The initial function of XCI may have been to reinforce differences in dosage of Sox3/Sry of XX and XY individuals and ensure that they develop into females and males, respectively. Intriguingly, two key genes for the evolution of XCI, Rnf12 and Lnx3, were probably physically close to Sox3 in the proto-X chromosome ~180 MYA, as suggested by the analysis of the location of these genes in mammals and birds [12]. Both the evolution of sex determination and XCI apparently involved the same relatively small region on the proto-sex chromosomes. The mechanism of XCI at the time must have been different from what it is today, as some key players such as Xist and Rex1 evolved later in the placental lineage [16, 21]. However, Lnx3 is involved in the ubiquitin pathway as is Rnf12 (see Introduction), which suggests that this pathway may have had a critical role in the early evolution of XCI. This hypothesis predicts that if XCI has evolved to silence Sox3 in females, Sox3 must be among the X-inactivated genes, which indeed appears to be the case in mice [54, 55]. At this point, this hypothesis is very speculative, but it has an interesting implication. Theory predicts that the early suppression of recombination between proto-X and proto-Y chromosomes must involve either at least two sex-determining genes [56], or a sex-determining gene and at least one sexually antagonistic (beneficial for one sex, harmful for the other) gene [57]. Selection will then favor suppressed recombination between sex-determining genes so that the male-determining alleles remain linked on the Y, and the female-determining alleles remain linked on the X. When a male-beneficial-female-detrimental gene appears on the sex chromosomes, selection will also favor it to be genetically linked to the male-determining locus [58]. In mammals, only one sex-determining gene, Sry, has been described, and it has been suggested that sexually antagonistic mutations may have accumulated in the vicinity of Sry very early in the evolution of the mammalian sex chromosomes, driving suppression of recombination between the X and Y [57]. If genes other than Sry were involved in the early evolution of sex determination of mammals as we suggest, this might be sufficient to explain early suppression of X–Y recombination, and sexually antagonistic genes might only have had a role in later in further suppressing recombination along the sex chromosomes [59]. However, more work is needed to test this hypothesis of multiple sex-determining genes in mammals.

The evolution of random XCI

XCI is found both in placentals and marsupials, consistent with an early evolution of this mechanism soon after the therian sex chromosomes originated ~180 MY ago. However, the mechanisms of XCI are different in both lineages [7–9]. The coating of the future inactivated X is mediated by a long non-coding RNA in both placentals and marsupials, but in placentals this RNA is coded by Xist (a gene that evolved specifically in the placental lineage from a protein coding gene; [21]) whereas in marsupials this RNA is coded by the non-orthologous gene Rsx [60]. This may indicate a parallel refinement of the mechanism for XCI independently in both lineages, or simply that XCI originated twice. Moreover, in placentals, XCI is random (rXCI), i.e. one of the two Xs is randomly inactivated in different cells during development (stage E.4.5 in mice when cells start to differentiate), which results in a mosaic of cells with differently inactivated Xs in somatic tissues. In marsupials, XCI always affects the X chromosome transmitted by the father, and is called paternal XCI (pXCI) [61]. In placentals, pXCI has been reported in extraembryonic tissues, and in embryonic tissue early in development prior to rXCI (reviewed in [13–15]). There are differences however among placentals. In mouse and cattle, pXCI has been observed in trophoblast tissues [13]. In other placentals studied thus far, no pXCI in has been reported in the extraembryonic tissues (human [62], rhesus macaque [63], rabbit [62], and horse [64]). A biased inactivation (inactivation of the paternal X was found more frequent than that of the maternal X) has also been reported in neonatal brain in mice, but the bias was small and interpreted as a residual of pXCI [65].

Explaining these differences in XCI between species is challenging. The current view is that pXCI is the ancestral form (Fig. 3; see also above), but why pXCI would have evolved first is not clear. One reason could be that evolving pXCI is easier from a mechanistic point of view. In line with this, pXCI appears to require fewer cis and trans factors than rXCI (a 250 Kb Xist-transgene is enough to recapitulate pXCI, where rXCI requires a 460 Kb one [15]). In particular, rXCI requires a counting mechanism to inactivate only one X and not both Xs, which is mediated by Rnf12/Rex1 [16]. pXCI does not require a counting mechanism as it is always the X from the father that is inactivated [7], and consistently Rex1 is absent in marsupials [16]. As argued in the previous section, according to the parental antagonism model of XCI there are also evolutionary reasons why pXCI is expected to evolve first.

Understanding why rXCI would evolve to supersede pXCI is also challenging. With pXCI, females are effectively haploid for the X chromosome and recessive deleterious mutations on the X will be expressed. rXCI will generate tissues made of mosaics of cells in which, overall, both alleles will be expressed. An obvious consequence of rXCI is thus restoring partially diploidy for the X chromosome. The typical example is red-green color blindness in humans. Red-green color blindness is due to a recessive deleterious mutation on an X-linked opsin gene. This condition affects mostly males as they will always express the deleterious mutation if present on their single X chromosome. Only females homozygous for this deleterious mutation in the X-linked opsin gene will be in a similar situation. In heterozygous females, the X-linked opsin gene is randomly X-inactivated. A sufficient number of cone cells express the functional allele and can sense color so that most heterozygous females will not be color-blind, which explains why red-green color blindness is much more common in men than women. Recent theoretical work has explored the conditions under which rXCI and pXCI may evolve [66]. If the alleles deleterious for female fitness are mostly recessive then rXCI is expected to evolve. If many sexually antagonistic alleles (beneficial for one sex, harmful for the other) are segregating in the population, rXCI is not expected to evolve [66]. For instance, alleles beneficial for males and deleterious for females will generate selection for pXCI [66]. However, it is not clear why paternal XCI should evolve rather than maternal XCI. A higher mutation rate in males than in females (male-biased mutation) would imply that paternally-inherited X chromosome will carry more deleterious mutations, and pXCI should be selected in that case [66], although the effect might be too weak and needs to be studied in more details. A stronger sexual selection in males will also favor pXCI, as the paternally-inherited X chromosome will tend to include sexually-antagonistic genes harmful to females [66]. Connallon and Clark claimed that sexual dimorphism and potentially sexual selection might be overall stronger in marsupials than in placentals, which this may explain why pXCI has been maintained in the former ([66] and references therein). However, if the strength of sexual selection is the only driver of the transition from pXCI to rXCI, it is difficult to understand why it has not evolved in strongly sexually dimorphic placental species, such as cervids, pinnipeds and some primates. Alternatively, the parental antagonism model states that pXCI in some placentals may have been conserved or re-evolved as this tissue directly mediates demand from the offspring to the mother [50]. The intensity of parental conflicts for offspring demand may explain why some species have pXCI and others do not but this is purely speculative [50]. Differences in the timing of zygote genome activation (ZGA), the process by which the genome of the zygote starts being expressed, may also explain why pXCI is found in some placentals and not others. In mice, ZGA occurs early and may require a fast way of achieving XCI, hence the presence of pXCI in early mouse development. In other placentals (e.g., humans, rabbits), ZGA occurs later and there is sufficient time for rXCI to be established [62]. However, the timing of ZGA does not correlate very well with the presence/absence of pXCI in placentals, although XCI has been studied in only a few placentals thus far [13].

Concluding remarks and future directions

During the last 50 years, Ohno’s hypothesis was taken for granted as it was difficult to see how XCI would have evolved otherwise. NGS data have completely changed our view of the evolution of XCI and dosage compensation in mammals, although the situation may again change when more data become available. Global dosage compensation that was once thought to be a paradigm seems now an exception rather than a rule, as many cases of partial dosage compensation have been reported in birds, fish, and some invertebrates (reviewed in [43]). Global dosage compensation is only confirmed for a handful of species such as Drosophila and C. elegans. If the evolution of dosage compensation is driven by dosage-sensitive genes as suggested by the results in mammals, we expect it to be partial and affect only a few genes on the X or Z chromosomes. The number of dosage-sensitive genes may also vary from one organism to another, and in species with many such genes on the sex chromosomes, a global dosage compensation mechanism may evolve. In species with large effective population size (Ne) such as Drosophila and C. elegans, selection is very efficient so that the number of genes in the genome that are effectively dosage-sensitive is increased. Such a relationship between Ne and the extent of dosage compensation remains to be investigated. Also, identifying dosage-sensitive genes in many species will be crucial to test many of the ideas outlined here. Mammalian species with new sex chromosome pairs (as in some rodents, see [67] and references therein) may be particularly interesting as they raise the question of how the dosage-sensitive genes on the former X cope without XCI, if they indeed lose XCI. Understanding why XCI has evolved in the first place is still a great challenge, as testing the different available hypotheses is not easy. Theoretical work is certainly needed to explore further the different hypotheses. In particular, the parental antagonism model has not been modeled formally yet. Understanding the transition from pXCI to rXCI will probably require systemic surveys of XCI in mammals, as only a few species have been studied thus far. Approaches using NGS may facilitate this task and provide a broad picture of the evolution of XCI in mammals in the near future.

Abbreviations

XCI

X-chromosome inactivation

PAR

Pseudoautosomal region

rXCI

Random X-chromosome inactivation

pXCI

Paternal X-chromosome inactivation

Xi

Inactivated X chromosome

PolII

RNA polymerase II

ZGA

Zygote genome activation

NGS

Next-generation sequencing

Ne

Effective population size

PAM

Parental antagonism model

XIC

X-inactivation center

Box 1: testing Ohno’s hypothesis with expression data

Ohno’s hypothesis has been tested by comparing the expression of the X chromosome to that of the autosomes taken together, the X:AA ratio.

Microarray versus RNAseq data

Initially, microarray data have been used for this test [22–26]. Microarray may give less precise estimates of expression levels [27]. Moreover, microarray data have to be filtered prior to analysis. The procedures for data filtering rely on arbitrary thresholds, which applied similarly to the X and autosomes remove many lowly expressed X-linked genes and generate an artifactual X:AA of 1 [68]. RNAseq data are supposed to give more precise estimates of expression levels. However, there is also some noise in RNAseq data due to unspecific mapping of RNAseq reads onto the genome, and how this noise is removed can also affect the results [69]. Removing this noise is at the heart of the controversy between the different studies using RNAseq [27–30, 35]. As the threshold for considering a given expression level different from 0 increases, the X:AA ratio increases and reaches a plateau at 1 [28]. It is clear, however, that when using conservative thresholds, the number of X-linked genes analyzed becomes small, and one cannot conclude from this about a “global” X hyperactivation [32].

X:AA, X:XX, and other expression ratios

Using X:AA expression relies on the assumption that expression were similar between the proto-sex chromosomes and the autosomes (XX:AA = 1). Using the present-day and ancestral expression of the X chromosome, the X:XX ratio, is thus a more direct way to test for Ohno’s hypothesis. Computing the X:XX ratio in mammals has implied finding an outgroup where the 1-to-1 orthologs of the X-linked genes are autosomal, namely birds [33]. This guarantees that only genes that were originally on the sex chromosomes before they diverged are analyzed, which is what should be done as dosage compensation is expected for these genes only. The new genes that evolved (e.g. through intra-X duplication or translocation to the X) after X and Y stopped recombining and diverged should not be included in studies on dosage compensation, are correctly excluded of the X:XX analysis but not in the X:AA ones. However, finding 1-to-1 orthologs between distantly related species may be difficult and result in a small number of genes being analyzed. Moreover, all these chromosome-wide comparisons may be problematic as different selective forces (dosage compensation, sexual selection) may affect expression levels [43, 70]. A more precise way of testing Ohno’s hypothesis is to study X-linked and autosomal genes that are expected to interact in some ways and for which equal dosage may be required. Considering genes from the same network is one possibility [33], and considering genes belonging to protein complexes is another [32].

Box 2: The parental antagonism model of X chromosome inactivation

The parental antagonism model (PAM) for the evolution of XCI was proposed by Haig [49, 50]. It is embedded within the general evolutionary theory of parental investment in offspring [71] and closely related to the kinship theory of genomic imprinting [49, 72]. The argument can be presented in a number of steps.

Step 0 A prerequisite for PAM to work is that offspring are provisioned with an adjustable amount of resources from their mother following fertilization. This is indeed the case in therians where resources are provided through the placenta during embryonic development.

Step 1 At the core of PAM is the expectation that there will be an evolutionary conflict between maternally and paternally derived genes within a developing organism with respect to the amount of resources provided by the mother of that individual. Both genes derived from the mother and from the father will be selected to induce the mother to provide resources. However, the optimal amount of resources provided may be greater for paternally than for maternally derived genes. This is because when females mate with multiple males during their lifetimes, paternal interests will be limited to the current offspring whereas maternal interests extend to all future offspring that a mother will have.

Step 2 The X chromosome is two-thirds of the time inherited from the mother but only one-third of the time from the father (simply because females have two Xs and males just one). As a consequence, genes on the X chromosome are expected to reflect maternal interests more than paternal ones. In particular, it is expected that genes coding for embryonic growth inhibitors will accumulate on the X chromosome, whereas growth enhancers will be scarce [51].

Step 3 As an evolutionary response to this accumulation of growth inhibitor genes on the X chromosome, there will be selection on paternally inherited genes on the X chromosome to inactivate these genes in embryos, thereby increasing embryo growth. This inactivation may then also spread to other genes on the paternally derived X, either for mechanistic reasons or for dosage compensation. The resulting state of inactivation of the paternally derived X (pXCI) is found in marsupials.

Step 4 pXCI entails that an organism becomes functionally haploid, so that recessive deleterious mutations on the maternally derived X chromosome will be expressed and reduce fitness. This may create selection pressure for random XCI (rXCI), alleviating this burden because half of the cells will then express the functional gene copy [50]. This transition from pXCI to rXCI does not involve parental conflict because the choice of which X chromosome is inactivated does not affect gene dosage.

Step 5 Nevertheless, parental conflict over which of the X chromosomes is inactivated may persist or re-emerge. This is because there may still be imprinted growth inhibitor genes on the X chromosome that are silenced when paternally inherited, so that the maternally derived X chromosome will be under selection to remain the active X. As a consequence, pXCI can re-evolve from rXCI, which may explain pXCI in mouse trophoblast tissues.