Mammalian X chromosome inactivation evolved as a dosage-compensation mechanism for dosage-sensitive genes on the X chromosome

Abstract

How and why female somatic X-chromosome inactivation (XCI) evolved in mammals remains poorly understood. It has been proposed that XCI is a dosage-compensation mechanism that evolved to equalize expression levels of X-linked genes in females (2X) and males (1X), with a prior twofold increase in expression of X-linked genes in both sexes (“Ohno's hypothesis”). Whereas the parity of X chromosome expression between the sexes has been clearly demonstrated, tests for the doubling of expression levels globally along the X chromosome have returned contradictory results. However, changes in gene dosage during sex-chromosome evolution are not expected to impact on all genes equally, and should have greater consequences for dosage-sensitive genes. We show that, for genes encoding components of large protein complexes (≥ 7 members)—a class of genes that is expected to be dosage-sensitive—expression of X-linked genes is similar to that of autosomal genes within the complex. These data support Ohno's hypothesis that XCI acts as a dosage-compensation mechanism, and allow us to refine Ohno's model of XCI evolution. We also explore the contribution of dosage-sensitive genes to X aneuploidy phenotypes in humans, such as Turner (X0) and Klinefelter (XXY) syndromes. X aneuploidy in humans is common and is known to have mild effects because most of the supernumerary X genes are inactivated and not affected by aneuploidy. Only genes escaping XCI experience dosage changes in X-aneuploidy patients. We combined data on dosage sensitivity and XCI to compute a list of candidate genes for X-aneuploidy syndromes.

The sex chromosomes of therian mammals (placentals and marsupials) originated from a pair of autosomes about 150 million years ago (1–5). The X and Y chromosomes gradually diverged after several events of recombination suppression, probably inversions on the Y chromosome (3, 6, 7). With the exception of two very small pseudoautosomal regions (PARs), the Y chromosome never recombines. Because of its nonrecombining nature, the Y chromosome has degenerated and lost most of its genes (reviewed in ref. 8). In contrast, during therian evolution the recombining X chromosome has retained many ancestral genes (6), gained new genes, and evolved new expression patterns for some genes (4, 9, 10).

Early in the differentiation of the sex chromosomes, most ancestral genes were present on both X and Y and the imbalance of gene products between males and females would have been small. As the attrition of Y chromosome genes progressed, an increasing number of loci were uniquely present on the X chromosome, implying a twofold reduction of their expression in males (XY) compared with females (XX). X chromosome inactivation (XCI) in females makes expression of X-linked genes similar in males and females (11). However, instead of solving the problem of dosage imbalance between autosomal and X genes, XCI seemed to expand it to females. Ohno proposed that the global expression of the X chromosome must have doubled in both sexes during evolution, solving the X:autosome imbalance in males, and suggested that XCI had evolved subsequently to reduce the output of X-linked genes back to the ancestral levels in females (1). Both steps are required to call XCI a dosage-compensation mechanism.

Consistent with Ohno's hypothesis, microarray data suggested that the mammalian X chromosome global expression level was similar to that of autosomes (12–14). However, analysis of RNA-seq data, which yield much more precise expression-level estimates than microarray data, indicated that X chromosome global expression level in humans and mice was half that of autosomes in both sexes (15). This analysis suggested that the first step in Ohno's scenario was missing and raised doubt about XCI as a dosage-compensation mechanism (16). It was recently shown that this conclusion was strongly affected by the inclusion of testis-specific X genes in the analysis (17) (see also refs. 18–22 about the controversy regarding dosage compensation in mammals). However, even though excluding these genes with no expression in somatic tissues brings X chromosome global expression closer to that of autosomes, it is still significantly lower, suggesting that the true nature of X dosage compensation may differ from these “all or nothing” scenarios.

In zebra finch, chicken, and crow, partial dosage compensation has been observed on the avian Z chromosome (23–26), with only some Z-linked genes being dosage-compensated (27, 28). Partial Z chromosome dosage compensation has also been observed in silkworm (29, 30) and in the parasite Schistosoma mansoni (31). Studies on the platypus indicate incomplete X chromosome dosage compensation in monotremes (32, 33). Data from sticklebacks show that the X is more strongly expressed in females than in males (34), consistent with a lack of global sex-chromosome dosage compensation in this fish. All this work suggests that global dosage compensation might not be a general feature of sex chromosomes (35).

Changes in gene dosage during sex-chromosome evolution are only expected to affect dosage-sensitive genes (35), which could explain why partial dosage compensation has been observed when analyzing all X/Z genes combined together. In this study, we focused on dosage-sensitive genes in the human genome and tested for dosage compensation of these genes only. Early experiments comparing polyploids and aneuploids in plants have shown that imbalanced expression of dosage-sensitive genes can strongly impact the phenotype (36). It was later shown that in yeast, most dosage-sensitive genes are involved in protein complexes (37). Using experimental data from strains heterozygous for single-gene knockouts, Papp et al. (37) could indeed show that a strong decline in fitness is only observed for genes encoding proteins involved in complexes (hereafter named protein-complex genes), such as the ribosome. They also found that proteins from the same complexes tend to be coexpressed at very similar levels and tend to have the same number of copies. All these lines of evidence, and others, suggest that there are strong constraints on the stoichiometry of the members of a complex and that an imbalance in such stoichiometry can be deleterious (37).

This “balance hypothesis” has become very popular and has been repeatedly used to explain patterns of duplicate gene evolution in yeast, Arabidopsis, rice, composites, and Paramecium (reviewed in refs. 38 and 39). In these organisms, most duplicate genes maintained after whole-genome duplication (WGD) events over large evolutionary periods are involved in protein complexes (40–45). In contrast, protein-complex genes are underrepresented in duplicates from segmental duplications (46, 47). These patterns of gene duplicability are fully predicted by the balance hypothesis because WGD events will not affect the stoichiometry of the components of a complex (but subsequent loss of a single component will be counter-selected), whereas segmental duplications will disrupt the stoichiometry.

It has been suggested that in multicellular organisms, selection for balanced dosage may be weaker than in unicellular organisms because selection is reduced in such organisms with small effective population size (48, 49). Additionally, genes involved in regulatory networks (such as transcription factors) are also expected to be dosage-sensitive, and in multicellulars these genes are probably numerous (39). However, in multicellulars, many dosage-sensitive genes are likely to be protein-complex genes. This idea was explored in humans and dosage-sensitive genes were identified as genes maintained after WGD events and resistant to segmental duplications and copy-number variations, and called dosage-balanced ohnologs (DBOs) (50). Protein-complex genes were found overrepresented among these DBO genes. Strikingly, 75% of the Down syndrome (trisomy 21) candidate genes are DBOs and a highly significant excess of DBOs was found in the Down syndrome critical region, which is known as a major determinant of the features of this syndrome. This finding is consistent with the observation that many haploinsufficient genetic diseases in humans are caused by protein-complex genes (51).

Here we focused on protein-complex genes in humans to test for the evolution of dosage compensation in dosage-sensitive X-linked genes. We used a list of protein-complex genes inferred from experimental data and expression-level estimates from RNA-seq data in humans. Based on these results we built a list of genes of interest for X aneuploidy syndromes, our rationale being that dosage-sensitive genes that escape X inactivation could be a cause of the phenotypes observed in these syndromes.

Results and Discussion

Expression Analysis of Dosage-Sensitive X Genes and Evidence for Dosage Compensation in Humans.

Assuming global autosomal expression level has not changed since the X and Y chromosomes originated, and if XCI has evolved to make autosomal and X expression equal (dosage compensation), as in Ohno's scenario, then the X/A mean expression ratio should be 1 in both males and females. A value of 0.5 in males is expected if X expression has remained constant along XY chromosome evolution, because males only have one copy of the X chromosome. In this case, a value of 0.5 is also expected in females because one of the two X chromosomes is inactivated and not transcribed. An X/A expression ratio of 0.5 would mean that XCI does not act as a dosage-compensation mechanism and its role is equivocal. Our independent analysis of data from 12 male and female tissues from ref. 15 found, as did the authors of the original analysis of that dataset, that X chromosome global expression is about half the expression of autosomes in both male and female (Fig. 1A). We checked whether differences in dataset/raw read processing could explain differences in conclusions found in Xiong et al. (15) and Deng et al. (17), but found expression-level estimates from both studies to be strongly correlated (Materials and Methods). Using data from Xiong et al. (15), we found that the X/A expression ratio significantly increases when the nonexpressed genes are removed, as in Deng et al. (17): it is now close to 0.7 for most of the tissues (Fig. 1A). Such an increase is explained by a higher fraction of nonexpressed genes on the X chromosome than on the autosomes; indeed, the X chromosome includes many testis-specific genes not expressed in somatic tissues in both humans and mice (17, 19, 22). Taken together, these data suggest that our work is not affected by differences in datasets or procedure for data-analysis, as we are able to make the same observations as in refs. 15 and 17 using different filters.

Fig. 1.

X/A expression ratio. (A) In this analysis, 734 X genes and 19,066 autosomal genes are included (Materials and Methods). Expression of X genes is normalized by the median of autosomal gene expression. The median of X/A ratios and associated 95% confidence interval are shown for each tissue. Results for both all genes (black, as in ref. 15) and excluding nonexpressed genes (gray, as in ref. 17) are shown. (B) Here only genes involved in protein complexes are included. For each complex, we computed the median of X gene expression over that of autosomal gene expression. We prepared three groups with similar sample size with increasing protein-complex size in number of proteins: small (2–3 proteins, yellow), medium (4–6 proteins, orange), large (7–120 proteins, brown). For each tissue and complex size category, the median of within-complex X/A ratios and associated 95% confidence interval are shown. In both panels we show the results for a pool of eight tissues (see text). The two green dashed lines indicate expectations with dosage compensation (X/A = 1) and without dosage compensation (X/A = 0.5).

Deng et al. (17) inferred that there is a global up-regulation of X gene expression in humans (see also ref. 19). For most tissues, however, the X/A expression ratio is close to 0.7 and is not 1, the expected value for global X up-regulation (Fig. 1A). Deng et al. (17) suggested that this is because RNA-seq data are noisy: genes with RNA-seq–estimated low expression levels can actually be nonexpressed genes (see also ref. 19). As these genes are comparatively more numerous on the X than on the autosomes, an X/A expression ratio smaller than 1 is expected from noisy RNA-seq data. Using brain as an example, Deng et al. (17) argued that when the expression-level distributions from X and autosomes are compared, they seem to be similar, which supports global X up-regulation (see figure 1A in ref. 17). However, many tissues do show nonoverlapping X and autosomal distributions (see supplementary figure 1 in ref. 17). Instead, we interpret these data as suggestive that the X chromosome includes a mixture of up-regulated and nonup-regulated genes, the combined analysis of which returns a “mixed” X/A expression ratio between 0.5 (no dosage compensation) and 1 (full dosage compensation).

To refine the analysis we examined genes involved in protein complexes from the Human Protein Reference Database (HPRD) list (Materials and Methods) because these genes are likely to be dosage-sensitive and should be the main target for dosage compensation. This list includes 207 human protein complexes with proteins from both X and autosomal genes. We computed the X/A expression ratio within each complex and we obtained the median of this ratio among all complexes present in a given tissue, as well as pooled for all tissues excluding “reproductive” organs (testis, breast) and brain/nervous system tissues, because X-linked genes are known to be overrepresented and overexpressed in these tissues (6, 15) (Fig. 1). A preliminary analysis showed that complex size is a strong determinant of the X/A expression ratio within protein complexes, as revealed by a multiple regression analysis, including complex size, whole-complex gene expression, and percentage of X proteins in a complex [effects on X/A from strongest to weakest: complex size, P value = 0.0007; percentage of X proteins, P value = 0.03; global complex expression, nonsignificant]. In Fig. 1B, we therefore showed the results for three protein-complex size categories, each containing the same number of complexes. This analysis clearly shows that the X/A expression ratio increases with protein-complex size. For large protein complexes (≥7 proteins), the X/A expression ratio is significantly higher than 0.5 for 11 of 12 tissues, as well as in the pooled expression data. For most tissues (10 + pooled expression data), the X/A expression ratio is not significantly different from 1, the value expected in case of dosage compensation. Our observation suggests that dosage-sensitivity is stronger for genes involved in large protein complexes than for genes involved in small protein complexes, which makes dosage compensation required more often for the former. Complex size has also been found to have some influence on dosage sensitivity in yeast, because the fitness effect of dosage imbalance in heterozygous knockout mutants is correlated to protein-complex size (37). Several explanations could account for this finding. First, if imbalance leads to incomplete (and nonfunctional) complexes that are destroyed by the cell, the bigger the complex, the bigger the metabolic cost for the cells. Second, subunits forming a bridge between parts of the complex can inhibit complex assembly if present in excess. This problem should increase with the number of subunits in a complex.

Importantly, our results are unaffected by inclusion or exclusion of nonexpressed genes (the patterns shown in Fig. 1B remain exactly the same after removal of nonexpressed genes) (Fig. S1). Deng et al. (17) showed that the X/A expression ratio increased from 0.5 to 1, also increasing the minimum expression-level threshold required for a gene to be included in the analysis (see figure 1D in ref. 17; see also ref. 19). These authors suggested this finding was because of noise in the RNA-seq data affecting lowly expressed genes (see above) and concluded that the X chromosome was probably up-regulated as a whole. Using a similar test, we found that when the X/A expression ratio reaches 1, only a small fraction of X genes (159; i.e., 21% of the initial set of X genes) are still being analyzed, which weakens the idea of global up-regulation on the X chromosome (Fig. S2). We also found that the fraction of protein-complex genes increases when applying different thresholds for gene expression and moving the X/A expression ratio from 0.5 to 1. Importantly, this pattern is stronger for large complexes than for other complexes (P value = 0.034, Fisher's exact test with two categories for expression level using data from ref. 15: ≤ 0.05 and > 0.05), which suggests highly expressed genes include more dosage-sensitive genes, a trend that has been noted before (52). When excluding lowly expressed genes from the dataset, we may be getting rid of noisy data but we also seem to be enriching the dataset in dosage-sensitive genes, which could help getting an X/A expression ratio of 1 in agreement with up-regulation of the X chromosome affecting mostly dosage-sensitive X genes.

A balanced X/A expression ratio for a given complex could result from a twofold increase of X gene expression (as in Ohno's scenario) to match autosomal gene expression, or a twofold reduction of autosomal gene expression to match X gene expression. To distinguish these possibilities, we split our dataset in two categories: large complexes (≥ 7 proteins; L) and other complexes (< 7 proteins; O) and computed the ratio of the expression in large complexes and in other complexes separately for X genes (X_L/X_O ratio) and for autosomal genes (A_L/A_O ratio). We found that the A_L/A_O ratio is close to 1 for all tissues (Fig. 2 and Fig. S1). This result indicates that expression of autosomal genes does not differ significantly between large and other complexes, so the dosage changes we observed (Fig. 1B) are not a general feature of high expression levels of large complexes; rather, this feature is restricted to the X chromosome. Only about 100 X genes are included in the computation of X_L/X_O ratio (Fig. 2 legend) and, as expected for such a small dataset, the error bars are large. Nevertheless, the X_L/X_O ratio is significantly higher than 1 for 9 of 12 tissues (P value = 4.9 × 10⁻⁴, Wilcoxon paired test) and in seven cases, this ratio is close to two, suggestive of a doubling of expression levels in agreement with dosage compensation. This doubling could be explained by an enrichment of RNA polymerase II in 5′ of the X genes compared with autosomal genes (17, 22) mediated by active histone marks (22).

Fig. 2.

X expression and autosomal expression in large protein complexes versus others. For each tissue, we computed the ratio of the median of X gene expression of large complexes (≥ 7 proteins, n = 59) and the median of X gene expression of other protein complexes (< 7 proteins, n = 52), which we called the X_L/X_O ratio (red). Both categories have been defined from results presented in Fig. 1B. The ratio of the median of autosomal gene expression of large complexes (n = 696) and the median of autosomal gene expression of other protein complexes (n = 151)—the A_L/A_O ratio (blue)—was computed similarly. Error bars have been obtained by bootstrapping protein complexes and computing both ratios and represent 95% bootstrap confidence interval. We pooled the data for eight tissues (see text) and computed the median and confidence interval the same way. The two green dashed lines indicate expectations with a twofold increase of expression (ratio of 2) and without any change in expression (ratio of 1).

We thus find evidence that the scenario put forward by Ohno of a twofold increase of X gene expression in both sexes plus inactivation of one X in female as a way of compensating for Y gene loss is valid for dosage-sensitive genes in humans. Our analysis focused on protein-complex genes, which are considered the main source of dosage-sensitive genes in yeast (37). In multicellulars, genes involved in regulatory networks may be another major source of dosage-sensitive genes (39). We know that the dosage of some X genes escaping XCI can modulate autosomal gene expression, although the effect is small (53). This finding suggests that many dosage-sensitive regulatory X genes may be compensated, and it would be interesting to test for dosage compensation in these genes.

Dosage-Sensitive XCI Escapees as Candidate Genes for X Aneuploidy Syndromes.

Most autosomal aneuploidies are nonviable, with the notable exception of chromosome 21. Interestingly, chromosome 21 is the human chromosome with the lowest number of dosage-sensitive genes, which suggests dosage-sensitive genes are key elements of aneuploidy phenotypes (50). The X aneuploidy in humans is known to have only mild effects, which at first sight may be surprising given the size and the number of genes of the X chromosome; aneuploidies of autosomes of equivalent size are all lethal. Sex-chromosome aneuploidies have a very high prevalence in humans, with Klinefleter (XXY) being the most common aneuploidy in men (1/500–600), Triple-X (XXX) being the most common in females (1/1,000), and Turner (X0) being quite common in females (1/2,000–2,500). This finding is explained by X-inactivation of all of the supernumerary X chromosomes, which means that in case of loss of one X or the presence of extra X chromosomes, only one X chromosome will be active, as in XX females (54, 55). Some genes, however, escape XCI and it has been proposed long ago that these genes could underlie Turner, Klinefelter, and other X aneuploidy syndromes (56–60). About 100 XCI escapees are currently known in humans from experiments on about 600 X genes (these are two-thirds of the X genes), which means that maybe about 150 X genes could escape XCI in total (58). This small number of genes could explain why X aneuploidies have an even milder effect than chromosome 21 trisomy (there are 223 genes on chromosome 21). Interestingly, in mice only 3% of X genes escape XCI, compared with 15% in humans, and X monosomy in mice has smaller phenotypic effects than in human, which is consistent with XCI escapees underlying X aneuploidy syndromes (61).

Dosage is clearly central in Klinefelter syndrome, as the neurodevelopmental and psychological features of patients become more severe as the number of supernumerary X chromosomes increases, for example in XXXY and XXXXY males (60). Very few candidate genes are known for any X aneuploidy syndrome. One well-established candidate gene is SHOX, a gene from PAR1 that is involved in small stature in Turner syndrome (62, 63). SHOX is haploinsufficient in Turner patients. In Klinefelter patients, SHOX escapes XCI and is overdosed and the prototypic Klinefelter patient is tall, which is consistent with SHOX being a Klinefelter gene (59, 60). The case of SHOX suggests that the same genes could underlie Klinefelter, Turner, and other X aneuploidy syndromes, which would make sense as these syndromes often relate to the same traits (e.g., stature, cognition). Another somewhat equivocal candidate for Turner syndrome is RPS4 (57, 63). RPS4 escapes XCI, has a functional Y homolog, and is located in Xq. This gene encodes a ribosomal protein and clearly falls in our dosage-sensitive gene category. Interestingly, it has been shown that 46,X,i(Xq) karyotype (i.e., isochromosome Xq) cannot be differentiated phenotypically from 45,X Turner syndrome patients (64). This finding was initially considered evidence that Turner syndrome genes are on Xp because Xp is missing in 46,X,i(Xq) patients. However, the 46,X,i(Xq) patients carry three copies of the RPS4X gene and the above-mentioned results are also consistent with overdosage of RPS4 being as deleterious as half-dosage, which fits well with the dosage-balance hypothesis. The case of RPS4 shows that dosage-sensitive genes may have similar phenotypic effects in Turner, Klinefelter, and other X aneuploidy syndromes, but other genes, such as SHOX, may have opposed phenotypic effects depending on gene dosage.

Our results on X chromosome protein-complex genes suggest that among the XCI escapees, those that are dosage-sensitive genes might have the strongest impact on the phenotype of X0, XXY, and XXX individuals. Importantly, these genes should impact X0, XXY, and XXX individuals in a similar way, as haploinsufficiency or doubled-dosage of protein-complex genes are expected to yield improper stoichiometry in both cases and be deleterious (37). We used the list of protein-complex genes on the X chromosome and identified those escaping XCI (Materials and Methods) as likely candidates for X aneuploidy syndromes (Table 1). This list includes the already known RPS4 Turner candidate gene. The most interesting candidates are probably those involved in large complexes because our results suggest that constraints on dosage are stronger for these. The persistence of a Y homolog also suggests strong constraints on dosage (65) and candidates with a Y homolog and involved in large complexes are in boldface in Table 1. This should not be considered an exhaustive list because the data on X-inactivation and protein-complex genes (and dosage-sensitive genes in general) are known to be partial.

Table 1.

List of candidate genes for X aneuploidy syndromes

Gene name*	Y homology	Max complex size†	Function annotation‡
PPP2R3B	Y homolog	3	Serine/threonine-protein phosphatase 2A regulatory subunit B
TBL1X	Y homolog	7	F-box-like protein involved in the recruitment of the ubiquitin/19S proteasome complex to nuclear receptor-regulated transcription units
RBBP7	-	16	Core histone-binding subunit, Component of several complexes which regulate chromatin metabolism
EIF1AX	Y homolog	3	Eukaryotic translation initiation factor 1A
SH3KBP1	-	7	SH3 domain-containing kinase-binding protein 1
USP9X	Y homolog	20	Ubiquitin-specific-processing protease FAF-X.
MED14	Y pseudogene	29	Mediator complex, a coactivator involved in the regulated transcription of nearly all RNA polymerase II-dependent genes
UBA1	Y homolog	40	Ubiquitin-activating enzyme E1
WAS	-	12	Effector protein for Rho-type GTPases. Regulates actin filament reorganization via its interaction with the Arp2/3 complex.
SMC1A	-	9	Central component of cohesin complex, required for the cohesion of sister chromatids after DNA replication.
RPS4	Y homolog	40	40S ribosomal protein S4, structural constituent of ribosome
MAGEE1	-	4	Hepatocellular carcinoma-associated protein 1
CHM	-	3	Rab proteins geranylgeranyltransferase component A 1
MORF4L2	-	27	Component of the NuA4 histone acetyltransferase complex
TRPC5	-	5	Transient receptor potential Ca2+ channel
PLS3	-	3	Actin-bundling protein
CUL4B	-	7	Core component of multiple cullin-RING-based E3 ubiquitin-protein ligase complexes
HCFC1	-	13	Host cell factor 1

*The best candidates (members of large complexes and with a Y homolog) are shown in bold.

^†Most of the genes are involved in several complexes in the list from HPRD (see Material and Methods), only the size of the largest complex is indicated here.

^‡From NextProt, the new database on human proteins developed by Swissprot (www.nextprot.org). Ubiquitin related genes are in bold.

Klinefelter syndrome is characterized by high stature, sparse body hair, gynecomastia, infertility, small testes, decreased verbal intelligence, and increased risks for autoimmune diseases (60). Features of Triple-X syndrome include tall stature, epicanthal folds, hypotonia, clinodactyly, seizures, renal and genitourinary abnormalities, premature ovarian failure, motor and speech delays, and increased risks of cognitive deficits and learning disabilities (59). Turner syndrome is characterized by short stature, premature ovarian failure, and a variety of anatomic abnormalities, including webbing of the neck, lymphedema, aortic coarctation, autoimmune diseases, and characteristic neurocognitive deficits (impaired visual-spatial and visual-perceptual abilities, motor function, nonverbal memory, executive function and attentional abilities; see ref. 63). Interestingly, some of the candidate genes have annotations reminiscent of these X aneuploidy features, although the syndromes are not explicitly cited (Table S1). In addition, three of the four best candidates (in large complexes and with a Y homolog; in boldface in Table 1) are involved in the ubiquitin pathway, which relates to protein degradation and addressing in the cell. Ubiquitination occurs in a wide range of cellular processes, such as differentiation and development, immune response and inflammation, neural and muscular degeneration, morphogenesis of neural networks, and ribosome biogenesis.

Conclusions

Our results open perspectives for finding candidate genes for X aneuploidy syndromes. Such syndromes are very common (up to 1 in 500 male births for Klinefelter) and, although the phenotypic consequences are mild and vary a lot among individuals, with some individuals being asymptomatic, many practicians call for efficient diagnosis because many X aneuploidy individuals can experience health, fertility, and cognitive difficulties if not treated (59, 60, 63). Surprisingly for such common diseases, very little is known about the genotype-phenotype relationships. We suggest dosage-sensitive genes that escape XCI should be tested, for example in animal models (66), as they seem to be good candidate genes for X aneuploidy syndromes.

Our results also show that Ohno's idea of a two-step dosage-compensation mechanism (twofold increase of X expression in both sexes plus an XCI in females) is valid for dosage-sensitive genes (i.e., protein-complex genes). How this two-step dosage-compensation mechanism evolved still needs to be understood. In Ohno's logic, the doubling step should come first and then the halving one (through XCI). However, we know that XCI is very old because the Xist locus is located within the earliest diverging segment of the sex chromosomes (stratum 1; see ref. 3) and XCI is found in both marsupials and placentals, which suggests XCI may have evolved first. As the range of XCI silencing crept along the chromosome, then X-linked dosage-sensitive genes (but not other genes) would have experienced selection for doubling of expression. However, in this case, the reason why XCI would evolve first is not clear. In marsupials and in some tissues (placenta, brain) of some placentals, XCI always affects the paternal X (67–71). Some authors suggested that XCI may have originally been a form of genomic imprinting related to parental conflicts (72, 73). Further work is needed to distinguish these two alternatives, but in any case, our work establishes the role of XCI in balancing expression between X and autosomal genes that are dosage-sensitive.

Materials and Methods

Expression Data.

We used gene-expression levels obtained from RNA-Seq data of 19,800 human genes (19,066 autosomal and 734 X) in 12 male and female tissues compiled by Xiong et al. (15). Sources of RNA-Seq data and methods are described in ref. 15 but, briefly, only reads uniquely mapped to exons were considered valid hits and expression level of a gene was defined by the number of valid hits to the gene divided by the effective length of the gene. For comparisons between tissues or developmental stages, expression levels were normalized by dividing the total number of valid hits in the sample. Genes with effective length smaller than 100 were discarded, resulting in 19,800 genes.

We cross-linked ref. 15 and ref. 17 datasets using gene names (as no other identifier was available in the latter). We could keep 9,835 genes, which revealed that expression estimates from both studies are strongly correlated: lung (Spearman ρ = 0.87), adipose (ρ = 0.91), brain (ρ = 0.92), colon (ρ = 0.88), heart (ρ = 0.94), liver (ρ = 0.94), lymph node (ρ = 0.90), muscle (ρ = 0.94), testes (ρ = 0.85), kidney (ρ = 0.89), breast (ρ = 0.88); all with a P value < 10–5.

Protein-Complex Data.

We obtained a list of members of human protein complex from HPRD release 9 (www.hprd.org). This list includes 1,521 annotated (and experimentally confirmed) protein complexes (74). Human genes and their chromosomal locations (X, autosomal) as described in Ensembl release 52 (www.ensembl.org) were assigned to the members of protein complexes using Ensembl IDs in HPRD. Using protein-complex IDs, we counted the number of the members for each complex to get the protein-complex size. Members without any Ensembl gene IDs were excluded, as well as complexes including only X or autosomal genes. This process led to a dataset of 207 complexes with proteins from 235 X and 1,381 autosomal genes and 89 X and 800 autosomal unique genes, as some genes are involved in several complexes.

X-Inactivation Data.

We used data on XCI from ref. 58. These data were obtained constructing nine different rodent/human somatic cell hybrids that retained an inactivated human X. National Center for Biotechnology Information (NCBI) build 34.3 annotations of X genes was used to design primers to amplify mRNAs and quantify X-inactivation of human X genes (58). Using these data, we classified as “inactivated” the genes that were significantly expressed only in two cells or less, as “escaping” the ones for which at least seven cells with a significant expression was observed, and as “heterogeneous” all other genes.

We checked all of the primers by blasting them on the updated X chromosome sequence from Ensembl release 60 (www.biomart.org). From the original 634 genes studied by Carrel and Willard (58), only 495 had both primers that matched both opposite strands and were separated by less than 100 Kb on the Ensembl release 60 X chromosome sequence. In some cases, several genes fell in the interval amplified by the same pair of primers; 69 genes were concerned. We excluded pseudogenes and selected the same gene as in ref. 58 when possible, and picked a gene at random in the interval otherwise. We also checked if primers matched on human autosomal chromosomes or on mouse X chromosome. Five genes had both primers that matched on human autosomes (EIF2S3, TIMM8A, SEDL, DDX3X, GLUD1) and four genes on the mouse X chromosome (DUSP21, HNRPH2, PHF16, ABCB7), and all were withdrawn to avoid false-positives of the RT-PCR experiment. We finally obtained a list of 392 genes. Among these, 55 are escapees, 304 are X-inactivated, and 33 are heterogeneous.

Statistical Analysis.

About 15% of X genes are known to escape XCI (58); we did not exclude these genes from the dataset as in ref. 15. Details on analysis are found in the figure legends. All statistical analyses were done using R.