Background

During early female development, one of the two X-chromosomes in each embryonic cell is randomly inactivated (Xi), leading to the formation of Barr bodies, which are packed corpuscles in the cell nuclei that are visible under a microscope1. Chromosome X inactivation (XCI) thus achieves gene dosage compensation with males, who carry only one X-chromosome per cell2 and is a mechanism to avoid lethal consequences of X-chromosome aneuploidies. However, not all the genes in Xi undergo inactivation. Under normal conditions, some genes always escape XCI, whereas others appear to escape depending on the tissue3,4. It is estimated that up to 23% of X-linked genes can naturally escape inactivation4,5. In contrast, in some aggressive breast tumor cells, the entire Xi is lost, as evidenced by the loss of Barr bodies6. In certain cases, this Xi loss is accompanied by a duplication of the activated X-chromosome (Xa), leading to increased expression of X-linked genes7. Loss of XCI has been observed in other cancer cells; however, the extent to which complete or partial XCI loss is a feature of female cancer remains unknown8. Given that XCI levels vary among genes, cells9 and tissues4, it is likely that global XCI maintenance also varies among women. However, the lack of a quantitative measure that can assess the level of defective XCI maintenance in a bulk biological sample remains a significant barrier to evaluating the health risks associated with XCI loss in women.

XCI occurs early in development and is considered to be stably maintained throughout the cells and lifespan of females10. Facultative heterochromatin is regulated in Xi by XIST, which has important consequences for development and disease11. Histone modifications and DNA methylation are the main epigenetic modifications on Xi12. DNA methylation is a key factor for XCI maintenance6. In this study, we propose that a consequence of defective XCI is the loss of methylation of consistently silenced genes3,5. As such, we computed the percentage of 2-allele demethylation of CpGs in XCI regions as a metric of defective XCI frequency in a sample and called it X-chromosome reactivation (X- Ra) level. Using data from The Cancer Genome Atlas (TCGA), we demonstrated the global transcriptomic consequences of X-Ra and determined its overarching role in female cancer. In breast cancer, we determined the molecular mechanisms likely driving the interindividual variations of X-Ra and validated its association with lower survival, tumor heterogeneity in triple-negative breast cancer (TNBC) and chemotherapy treatment, in independent studies. Finally, we used a large methylome study and independent studies to show that X-Ra was detectable in the blood and increased with age and cancer diagnosis. Our study thus proposes the first personal biomarker of defective XCI in women, and demonstrates its translation potential in the management and treatment strategies of female cancer.

Results

Study workflow

We initially observed that CpGs of X-linked genes such as AR, IL2RG and MECP2, which are widely recognized to be under consistent silencing, showed no 2-allele demethylation across healthy breast tissue samples from the TCGA. In contrast, known XCI escapees such as DDX3X, EIF1AX and KAL1 showed almost complete 2-allele demethylation (Fig. 1). We then hypothesized that 2- allele demethylation in CpG sites mapping to X-linked genes with reported consistent silencing in different studies4,5 would provide evidence for their defective control or maintenance. Under normal conditions, the strong control of XCI over epigenetic modifications on the Xi chromosome would imply widespread methylation of CpGs over the entire silenced region of Xi13 and, consequently, absence of 2-allele demethylation events in the region (Fig. 2). We then reasoned that, in a bulk biological sample, the global frequency of defective XCI could be measured by the percentage of infrequent 2-allele demethylation events in CpGs within Xi regions with consistent reported evidence of X-linked gene silencing. We first aimed to show evidence that this is the case, we then aimed to demonstrate that the percentage of defective XCI in bulk data could be robustly detected in different tissues and correlated with relevant clinical cancer outcomes, using multiple lines of evidence from independent studies (Tables 1–2). The study design and workflow are shown in Fig. 2.

Fig. 1: Methylation beta level distribution of CpGs in X-linked genes.
Fig. 1: Methylation beta level distribution of CpGs in X-linked genes.
Full size image

Methylation levels of CpG sites across X-linked genes that are widely recognized for consistent silencing (XCI) or consistent escape from XCI (ESC) in healthy breast samples from the TCGA.

Fig. 2: Flowchart of the study.
Fig. 2: Flowchart of the study.
Full size image

A We determined that XCI exerts a strong control on the methylation levels in inactivated genes in Xi. 2-allele demethylation in both copies of CpGs is not observed in genes under XCI under strong XCI maintenance. 2-allele demethylation (from Xi and Xa) may be observed for a defective XCI level. B We define X-Ra as the percentage of CpGs in genes under XCI that are observed with 2-allele demethylation. Here, seven XCI CPGs are illustrated on the red band of Xi. In the study, we consider a total of 4862 CpGs. C As a marker of tumor stratification, we assess X-Ra correlations with transcriptomic data, DNA somatic mutations, cancer risk, survival, and recurrence. As a blood biomarker, we assess the correlations of X-Ra with aging and cancer. Microarray methylation data was analyzed for different studies shown in the boxes, the acronyms of the studies are the accession numbers of GEO and TCGA. Further details of the studies, data and analysis are found in Tables1–2 and “Methods”.

Table 1 Studies included in the analysis of X-Ra and its biological correlations
Full size table
Table 2 Characteristics of TruDiagnostic DNA Biobank
Full size table

X-Ra definition

We first selected CpGs that were mapped to X-linked genes with strong previously published evidence of consistent inactivation4,5. We then estimated the frequency of defective XCI in a female biological sample by surveying the number of CpGs that were observed with 2-allele demethylation, i.e., methylation beta values < 0.2. We defined X-Ra (X-Reactivation) as the percentage of these events, to measure the quantitative level of partial activation of Xi supported by methylation removal. Therefore, under a strong XCI, X-Ra is expected to be low, and when XCI is weakened, X-Ra should increase. A list of 4862 CpGs was compiled from harmonized CpGs mapped in TCGA workflow to genes reportedly under XCI4,5. We compiled a list of X-linked genes with consistent silencing using the lists from Balaton and Brown5 and Tukiainen et al.4 In the first case, we selected 131 genes that Balaton and Brown reported as consistently subject to XCI based on four different metrics. In the second case, we selected 431 genes identified by Tukiainen et al. as consistently inactive, based on their allele-specific expression analyses combined with two previous studies. From these, we derived a common list comprising 440 X-linked genes with strong evidence of being subject to XCI.

For each study, we selected CpGs in the reference panel and extracted the beta methylation levels. Finally, for each individual, X-Ra was the percentage of CpGs under XCI with beta values lower than 0.2 (See “Methods”).

X-Ra as a measure of a patient’s defective maintenance of XCI

Using methylation data of 446 constitutional biological samples from 12 different tissues from TCGA (Table 1), we established X-Ra as a measure of defective XCI frequency in a biological sample. We first observed that CpGs under XCI with 2-allele demethylation were infrequent and varied among women. Across all 12 healthy tissues, we found that 96% of the 2-allele demethylation calls were in 20% of women and 64% in only 5% of women (Fig. 3A).

Fig. 3: X-Ra as a measure of defective XCI in 12 undiseased tissues from 12 cancer studies of the TCGA.
Fig. 3: X-Ra as a measure of defective XCI in 12 undiseased tissues from 12 cancer studies of the TCGA.
Full size image

A Low population frequency of CpGs under XCI with beta methylation values < 0.2 (2-allele demethylation) in 12 constitutional healthy tissues of the TCGA (Table 1). B Enrichment of chromatin states and DNase hotspots by XCI CpGs with 2-allele demethylation, taking all XCI CpGs as reference. C The red horizontal line shows the mean X-Ra across women in healthy tissues of each TCGA study. In blue, the distribution of mean X-Ra for random permutations of the XCI levels of CpGs. D X-Ra computed for rare XCI CpG 2-allele methylation (frequency <0.05) Vs all X-Ra computes with all CpGs mapping to genes reportedly under XCI.

We then annotated XCI CpGs with 15-chromatin states in blood and DNase hypersensitivity regions in six constitutional tissues (breast, monocytes, gastric, ovary, muscle, small intestine)14 downloaded from the ROADMAP Epigenomics.

Mapping Consortium (ChromHMM v1.10)15 (https://egg2.wustl.edu/) with cell type accession numbers (E029, E094, E097, E100, E109)16. We observed that while XCI CpGs with 2-allele demethylation spread through the chromosome (Supplementary Fig. 1), they were significantly enriched in transcription start sites (TSS) and DNase hypersensitivity regions, in relation to all the CpGs under XCI. They were also significantly depleted in heterochromatin, ZNF genes, and repeats, weakly repressed PolyComb, and strong transcription states (Fig. 3B). These results were in line with previous work showing that the CpGs in the island-containing promoters of genes under XCI are consistently methylated in Xi3,17, and with the notion that the more accessible the CpGs are, the more likely they are to undergo 2-allele demethylation.

We validated that X-Ra, i.e., the proportion of 2-allele demethylation in XCI CpGs, was a measure of XCI status, as random permutations on XCI labels of the CpGs in X robustly showed higher mean values of X-Ra than correct labeling in 12 different tissues (P < 0.001 in all cases, Fig. 3C). We also observed that 67% (P = 7.12 × 10−28) of the X-Ra variance was explained by X-Ra computed with only the 2-allele methylation calls that were present in less than 5% of the women, demonstrating that the infrequent demethylation of CpGs in Xi is the main contributor of the measure (Fig. 3D).

We subsequently confirmed typically low values of X-Ra in females (Fig. 4A), suggesting high XCI maintenance in women, but not in men (Supplementary Fig. 2). We also confirmed the increased number of 2 allele-demethylation events in escapees (Fig. 4A) and in genes with reported variable XCI escape4 (Supplementary Fig. 3). Overall, our analyses showed that X-Ra can be regarded as a suitable and robust quantitative measure of global defective maintenance of XCI in women.

Fig. 4: Methylation beta level distribution of CpGs in inactive genes and escapees.
Fig. 4: Methylation beta level distribution of CpGs in inactive genes and escapees.
Full size image

A Distribution across 12 healthy tissues of TCGA samples. The red line represents the distribution of methylation levels for CpGs under XCI mapping to X-linked genes with consistent reported silencing by Balaton and Brown and Tukiainen et al. The blue line represents the distribution of methylation levels for CpGs of genes that consistently escape XCI in Tukiainen et al. 2-allele demethylation is a feature of escapees and not of inactivated genes (grey area under the blue curve). X-Ra is the area under the red curve (in percentage points). B Distribution across 12 cancerous tissues of TCGA samples. X-Ra substantially increases as the area under the red curve is larger for values < 0.2. The blue area under the blue line about 0.5 is the percentage of CpGs in escapees in the 1-allele methylation zone, suggesting escapee silencing.

X-Ra as a marker of female cancer

We examined the association of X-Ra with the tumor status of samples from the 12 previously studied tissues (Supplementary Data 1). Our analysis revealed that tumors exhibited a significant increase in the percentage of XCI CpGs with 2-allele demethylation, reflected in their elevated X-Ra values (Fig. 4B). To investigate possible molecular mechanisms associated with X-Ra in cancer, we performed a differential expression analysis for each tumor (Supplementary Fig. 4). We found several significant genes at the genome-wide level (P < 3.79  ×  10−6). We observed a small but significant and consistent reduction of XIST expression with X-Ra (Log2FC = −0.067, P = 1.32 × 10−11, Fig. 5A, Supplementary Data 2), indicating a potential underlying relationship between X-Ra and XIST-dependent silencing. Meta-analyses of 13,192 genes from cancer studies identified 537 genome-wide significant genes (Supplementary Data 2). Thirty-two of these genes have been reported to be under XCI (enrichment OR = 2.87, P = 6.5 × 10−7), underlying an expected consequence of partial activation of Xi. We also observed that differentially expressed genes were more frequent from chromosomes 20, 7 and X (Supplementary Fig. 5). The top genome-wide significant genes included LRCH1 (Leucine-Rich Repeats and Calponin Homology Domain Containing 1) and PHF11 (Plant Homeodomain Finger Protein 11), which play important roles in CD8 + T cell response against tumors and pathogens18, as well as T-cell activation in autoimmune disease19. Enrichment analyses of these significant genes revealed several pathways relevant to cancer, such as signal transduction, cell proliferation, apoptosis, and viral response (Supplementary Fig. 6). Overall, these results demonstrated a genome-wide transcriptomic signature associated with lower levels of X-linked DNA methylation, indicating unstable XCI and/or defects in XCI maintenance.

Fig. 5: X-Ra as a marker of multiple cancer types.
Fig. 5: X-Ra as a marker of multiple cancer types.
Full size image

A Meta-analysis for the differential expression of XIST by X-Ra across the 12 TCGA cancer types. Cancer sites included: breast (BRCA), bladder (BLCA), colon (COAD) head and neck (HNSC), kidney (KIRC), liver (LIHC), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), pancreas (PAAD), rectum (READ), thyroid (THCA) and uterus (UCEC). B Meta-analysis across 12 different cancer types from the TCGA for additive cancer risk effects of the percentage of CpGs under. XCI with 2-allele demethylation (in % units) and the percentage of CpGs in escapees with 1-allele methylation (in % units). C Probability of loss of function somatic mutations (LOF) in genes under XCI as a function of X-Ra (left). Probability of LOF in escapees as a function of the percentage of 1-allele methylation in escapees (right).

We observed that the cancer status of the biological samples was associated with high X-Ra levels, adjusted for age (OR = 1.19, P = 6.05 × 10−5, Supplementary Fig. 7). Heterogeneity suggests tumor stratification according to X-Ra. While demethylation across the entire X-chromosome is a general feature of cancer (Supplementary Fig. 8), the risk estimates of X-Ra across tumors could still be substantially explained by the reduction of X-Ra as a measure of XCI status in cancer (Pearson’s R = 0.66, P = 0.018). Furthermore, the percentage of 2-allele demethylated CpGs of escapees, i.e., not under XCI (grey area under the blue curve Fig. 4B), was not significantly associated with cancer status (OR = 0.97, P = 0.46) (Supplementary Fig. 9). These results suggested a limited role of overall demethylation of the X-chromosome in the association between X-Ra and cancer.

Importantly, the percentage of CpGs in escapees with methylation values in the 1-allele methylation zone (0.2 < beta < 0.6, blue area under the blue curve Fig. 4B) increased in tumors, suggesting escapee silencing or defective escape of these genes in cancer. The percentages of 2-allele demethylation in XCI CpGs (X-Ra) and 1-allele methylation in escapees were strong additive factors for cancer risk-adjusted for age (XCI CpGs OR = 1.45, P = 4.09 × 10−6, CpG in escapees OR = 1.90, P = 3.22 × 10−9) (Fig. 5B). Furthermore, X-Ra was significantly associated with a lower probability of LOF somatic mutations in genes under XCI (OR = 0.98, P = 9.71 × 10−5, Fig. 5C), adjusting for cancer type. In contrast, the percentage of 1-allele methylation in escapees was associated with a higher probability of LOF somatic mutations (OR = 1.04, P = 1.82 × 10−4, Fig. 5D), suggesting that the risk of cancer increases when both methylation and mutations affect both escapee alleles. The strong negative correlation between the measures of defective XCI and escapee silencing (Pearson’s R = −0.39, P = 9.72 × 10−138, Fig. S10) indicated underlying tumor stratification that suggests differences in the mis-regulation of X-linked genes from Xi.

To determine the risk of defective XCI maintenance in cancer associated with high X-Ra levels, we used a receiver operating curve (ROC) analysis. The best threshold (Youden’s index) was identified at an X-Ra level of 10%, which showed a specificity of 85.6%. The attributable risk of dichotomized X-Ra for cancer status was 40% (95%CI = 23–56%). The proportion of high X-Ra (>10%) was 43% across cancer types, with the highest rates in the thyroid (75%), liver (56%), and breast (50%), and the lowest in lung adenocarcinoma (13%) and pancreas (17%), in line with the association with cancer status (Fig. 4B).

X-Ra as a marker of breast cancer

We investigated the association between X-Ra and breast cancer (BRCA) subtypes, survival, chemotherapy treatment and somatic mutations. A 1%-point increase in X-Ra in tumor biopsy increased the hazard ratio of death from BRCA by 1.7%, adjusting for age (HR = 1.017, P = 0.025) (Fig. 6A). We downloaded tumor purity for BRCA and tested whether it could be a possible cofounder. We observed a strong association between X-Ra and tumor purity (β = 0.58, P = 1.24 × 10−27), adjusted for age (Supplementary Fig. 11). However, purity levels were not significantly associated with survival (HR = 1.005, P = 0.23), consistent with previous studies showing survival associations for purity only in specific cancers20. We then adjusted the effect of X-Ra on survival for purity, and observed a similar effect size and a loss of significance (HR = 1.014, P = 0.1), suggesting reduced power rather than confounding by purity. We also computed mean methylation beta values across the genome. We found a strong negative association with X-Ra (β = −2.15, P = 1.95 × 10−49) and a trend to significance with survival (HR = 2.28 × 10−4, P = 0.0502), adjusting for age (Supplementary Fig. 12). When we adjusted the effect of X-Ra on survival for global mean methylation, we observed again a similar effect size and a loss in significance (HR = 1.013, P = 0.15), indicating loss of power rather than confounding by global cancer demethylation.

Fig. 6: X-Ra as a marker of breast cancer (BRCA).
Fig. 6: X-Ra as a marker of breast cancer (BRCA).
Full size image

A Association between X-Ra and the survival time from breast cancer diagnosis to death (months) in TCGA and validated in GSE78754 for TNBC. B Predicted probability of TNBC by X-Ra in TCGA (TNBC Vs ER + /PR + /HER + ) and validated in GSE225845 (TNBC Vs HE2R + /HER + ). C Reduction of X-Ra by chemotherapy treatment for breast cancer in GSE207460 and validated for TNBC in GSE184159. D Genome-wide association for the frequency of copy number gains and deletions in BRCA with X-Ra, in TCGA, using sliding windows of 0.05MB and duplication frequency >0.01. The highest peaks are observed in the RTL1-TNFRSF6B and MYC loci for gains, and in CSMD1 for deletions. E Lack of association between X-Ra and GOX and significant association between X-Ra and UPD(X).

Adjusting for age, we validated the effect of X-Ra on survival in an independent study (GSE78754) of 70 patients with triple-negative breast cancer (TNBC) (HR = 1.033, P = 0.0081), a subtype of breast cancer characterized by the absence of estrogen (ER-), progesterone (PR-) and human epidermal growth factor 2 (HER2-) receptors. Consistent with these findings, we found in GSE141441 that

95 TNBC patients who did not undergo surgery had a higher probability of recurrence due to a high frequency of X-Ra, after adjusting for chemotherapy treatment (OR = 1.08, P = 0.04) (Supplementary Fig. 13).

With respect to breast cancer subtypes, we found that X-Ra significantly decreased the probability of having ER- (OR = 0.982, P = 0.004), PR- (OR = 0.984, P = 0.008), and HER2- status, although not significantly (OR = 0.988, P = 0.15). Consistent with these results, X-Ra was significantly associated with TNBC (OR = 1.029, P = 3.1 × 10−4). This finding was validated in 224 patients (GSE225845), where X-Ra was associated with TNBC versus HE2R + /HER+ status (OR = 1.07, 95% CI = 0.0015) (Fig. 6B). In summary, we observed that X-Ra stratified patients with breast cancer, particularly TNBC, and was associated with important clinical outcomes.

We analyzed longitudinal studies to determine whether X-Ra could be used as a monitoring biomarker of neoadjuvant chemotherapy for breast cancer. DNA methylation analysis was done in tumor biopsies taken at different time points during the treatment. In 290 patients with three visits during chemotherapy treatment (GSE207460), we found a significant reduction of 1.78% in X-Ra for every 12 weeks of treatment (β = −1.78, P = 1.78 × 10−13) (Fig. 6C), adjusting for whether patients were additionally treated with bevacizumab. We further adjusted for inferred tumor purity using InfiniumPurify. While purity was strongly associated with time of treatment (β = −0.057, P = 1.08 × 10−16), we found that the association between X-Ra and time of treatment remained significant after adjusting for tumor purity (β = −0.60, P = 0.007). Interestingly, we noted that patients treated with bevacizumab further reduced X-Ra at each 12-week visit (interaction β = −1.30, P = 0.013). We validated the reduction of X-Ra with neoadjuvant chemotherapy in 22 TNBC patients (GSE184159) with four cycles of chemotherapy during 12 weeks (β = −3.56, P = 0.014), suggesting a stronger modifiable nature of X-Ra by chemotherapy for this subtype.

X-Ra associations with somatic mutations in breast cancer

We then investigated the potential molecular mechanisms underlying X-Ra activity in breast cancer tumors. We considered chromosomal copy number gains in 0.5MB windows. We discovered 492 windows with significant increments in log(X-Ra) associated with the frequency of gains (Fig. 6D, Supplementary Data 3). Remarkably, most of the top associations on each chromosome included notable oncogenes (Supplementary Table 1). We observed gains in MYC locus21, which is known to induce pluripotent stem cell reprogramming with reactivated X-chromosomes in females22. The most significant association was found in a window that included TNFRSF6B, which prevents apoptosis23 and RTEL1 which encodes a regulator of telomere length24. X-Ra in BRCA was significantly associated with longer telomere length (β = 6.6, P = 0.003)25. These data underscore the potential mechanism of X-Ra in BRCA, consistent with the association between telomere length and RTEL1 gains (TL difference = 1.72, P = 6.47 × 10−5). Additionally, 22 significantly deleted regions were associated with X-Ra, all located in the short arm of chromosome 8, with the highest peak at CSMD1 (Fig. 6D), which encodes a complement inhibitor known to act as a tumor suppressor in breast cancer26. We performed association analyses of copy number alterations with global mean methylation values, to explore the extent to which global demethylation, expected in cancer, could explain the observed associations of X-Ra. Interestingly, we found limited evidence of gains and deletions associated with global mean methylation (Supplementary Fig. 14). Using allele-specific copy number segment data available for 1043 TCGA-BRGA samples (workflow ASCAT2), we detected 247 samples with complete gains of chromosome X (GOX) and 34 with uniparental isodisomy for the entire chromosome (UPD(X)). While did not find any association with somatic copy number loss, gain in any region or in the entire X-chromosome (β = −0.08, P = 0.14, Fig. 6E), we observed a significant association between X-Ra and UPD(X) (β = 0.97, P = 2.45 × 10−14); an isodisomy that involves entire Xi loss and Xa duplication, and a feature of basal-like breast cancers6.

The association of X-Ra with 18 frequent somatic mutations (>5%) revealed an increase of X-Ra by 3.42 percentage points for eachTP53 mutation (P = 5.42 × 10−5) and decreased by 5.42 for each CDH1 mutation (P = 4.93 × 10−6). Consequently, X-Ra is likely associated with the basal-like 1 subtype of TNBC27.

X-Ra in blood is associated with cancer and aging

We analyzed methylation data from three different studies to test the association of X-Ra with age >65 years (median age of cancer diagnosis), adjusting for immune cell count in the blood. We excluded patients with a cancer diagnosis in TruDiagnostic; cancer diagnosis was an exclusion criterion in the other two studies. We observed significant increases in X-Ra with age in all studies (β = 0.44, P = 3.3 × 10−3, Fig. 7A, B), with significant heterogeneity given by GENOA, a pedigree study on hypertension. Associations with continuous age were significant in TruDiagnostic (β = 0.014, P = 0.0038) and MESA (β = 0.029, P = 3.81 × 10−5), but not in GENOA (β = 0.0018, P = 0.65). In the TruDiagnostic data, the association between X-Ra and continuous age increased 3-fold in menopausal women (β = 0.044, P = 6.64 × 10−5, Supplementary Fig. 15). In addition, and in contrast to cancer tissue, X-Ra was associated with shorter telomere length (β = −1.008, P = 1.19 × 10−5, Fig. 7C) and with biomarkers of biological aging derived from DNA methylation patterns that do not include X-chromosome CpGs. These included Horvath’s age (β = 0.029, P = 4.85 × 10−6), a measure of chronological age, and Levine’s methylation age (β = 0.016, P = 0.006), which incorporates health-related metrics of aging. These findings suggest that X-Ra in the blood serves as an indicator of female advanced aging.

Fig. 7: Association of X-Ra in blood with aging and cancer.
Fig. 7: Association of X-Ra in blood with aging and cancer.
Full size image

A Meta- analysis for the association between X-Ra and age (>65 years) across three large methylation studies in adult women. X-Ra significantly increases in elder women in each of the studies. B X-Ra over 5-year age groups in the TruDiagnostic study (***<0.001, **<0.01, *<0.1). C Association of X-Ra with telomere length in the. TruDiagnostic study.

Differential expression analysis of X-Ra in monocytes from the MESA study revealed significant genes relevant to breast cancer and histone modifications (Supplementary Fig. 16), such as upregulation of ABTB228, downregulation of SOGA129,30, and upregulation of RBBP4 and RSPO231. Additionally, we observed significant enrichment of inactivated genes in the differential expression analysis of X-Ra (OR = 3.00, P = 0.002, Supplementary Table 2) and a nominal, yet expected, decrease in the expression of XIST (log2FC = −0.024, P = 0.02). These findings suggest that elevated X-Ra levels in the blood are associated with breast cancer-related genes and exhibit transcriptomic signatures consistent with XCI maintenance.

We found a significant association between X-Ra in the blood and cancer diagnosis in the TruDiagnostic cohort (167 were diagnosed with any type of cancer) and observed an increase of 7.9% in cancer diagnosis for a 1%-point increase in X-Ra (OR = 1.077, P = 0.034, Fig. 8A). We did not find significant associations between cancer and methylation ages of Horvath and Levine clocks. The association between X-Ra in the blood and cancer in GSE237036, adjusted for age and immune cell abundance, was not significant (OR = 1.54, P = 0.25). However, the combination of both studies showed a significant association (OR = 1.08, P = 0.02), with no apparent heterogeneity (P = 0.34). The association between X-Ra in the blood and cancer was independently validated in a longitudinal study (GSE142536) of 17 elderly women, where cancer diagnosis increased the frequency of X-Ra by 2.2% (β = 2.2, P = 0.006) (Fig. 8B). No association was observed between Levine’s methylation age and cancer (Supplementary Fig. 17). Taken as a test-retest study, X-Ra showed high statistical reliability (ICC = 0.95, 95%CI = 0.91 0.98).

Fig. 8: Association of X-Ra in blood with cancer.
Fig. 8: Association of X-Ra in blood with cancer.
Full size image

A Association of X-Ra in blood with a probability of having been diagnosed with cancer in the TruDiagnostic study. B Association of X-Ra in blood with cancer. A longitudinal study (GSE142536) of methylation changes in blood through major surgical intervention. Methylation was assessed one day before intervention (BL), one day post-operative intervention (PoD1), and 4–7 days post-operative intervention (PoD4/7). The overall effect of cancer is shown with 95% confidence bands from a mixed model with random effect on the subject and adjusted by age.

Discussion

Our study provides evidence that the frequency of defective XCI in bulk samples is a novel source of inter-individual variation among women. We introduced X-Ra, the first personal measure of defective XCI. This measure is based on the frequency of demethylation marks arising from the likely global weakening of XCI maintenance in an individual. Previous studies have used DNA methylation at CpG islands to predict the XCI status of X-linked genes, classifying each gene as either subject to XCI or escaping XCI3,5. Similarly, allele-specific transcription has been used to determine the XCI state of genes across different tissues4. However, these results had not been integrated into a global measure of defective XCI that could be used as a personal diagnostic or prognostic biomarker. Our biomarker X-Ra addresses this gap, enabling the assessment of risk estimates in large population cohorts associated with defective XCI.

We thus show that poor XCI maintenance appears to be a hallmark of female cancer in general7. While X-Ra is risk factor for women, we also observed that 1- allele methylation of escapees—associated with a higher frequency of LOF mutations—is an additional factor to female cancer. These results suggest a broader misregulation of X-linked genes from Xi in female cancers, characterized by the aberrant expression of normally silenced genes and/or the silencing of typical escapees.

The silencing of escapees is consistent with the cancer risk associated with mosaic loss of the Y chromosome (mLOY) in males. In male cancer, mLOY is linked to LOF mutations in escapees with homologous copies on the Y chromosome32 that indicate important loss of transcription of Y-linked genes and their homologous escapees. Our results, therefore, offer an integrated picture of gonosome dysregulation in cancer for both sexes. Consequently, defective XCI maintenance in silenced genes and disruption of escapees should be targeted in further studies to assess their potential as diagnostic or prognostic biomarkers in female cancer33.

Our blood analyses supported the notion that a possible mechanism of cancer risk from defective XCI is by altering the immune response34, in line with the immunosurveillance hypothesis of mLOY in males32,35,36. While mLOY increases the risk of cancer and overall mortality in males36, the assessment of gonosome contributions to the observed male-biased cancer risk across multiple cancer types37 must also account for defective XCI regulation and escape in females.

We observed that X-Ra was a powerful insightful marker for breast cancer. X-Ra was associated with moderate but highly significant and consistent reductions of XIST expression38, 20q (RTEL1), and 8q (MYC) copy number gains21; 8p (CSMD1) deletions26; lower breast cancer survival39 and basal type 1 TNBC27. In addition, X-Ra was responsive to chemotherapy. These results show that defective XCI maintenance is an important mechanism underlying tumor heterogeneity, particularly in TNBC. TNBC displays heightened aggressiveness, proliferation, and propensity for early metastasis compared with hormonal subtypes40. Treatment is complicated by the absence of defined molecular targets owing to its heterogeneous nature. Previous stratification of TNBC usually relies on substantial data reduction, population dependence, and needs mechanistic interpretation41,42. We showed that X-Ra reliably defines a new type of breast cancer that is associated with TNBC by a clear mechanism, opening the door to the design of novel intervention strategies based on XCI regulation.

As a developmental process, it is important to highlight the role of MYC in cancer and its regulation of pluripotency in embryonic stem cells43, which exhibit significant differences in the global regulation of XCI44. Here, we demonstrate that defective XCI maintenance is a feature of a population of tumor cells characterized by MYC gains, among others, as we analyze bulk tumor biopsies.

Additionally, we show that XCI maintenance is a characteristic of individuals. While the X-Ra score in tumor samples may warrant further investigation regarding its relationship with somatic mutations, such as UPDs, their effect in X-Ra in blood and constitutional samples is negligible. The prevalence of UPD is approximately 4 per 1000 in individuals over 75 years old. Therefore, from our study in blood, it is noteworthy to highlight the association between interindividual variation in defective XCI and aging. Consequently, it is important to further investigate modifiable and genetic factors that may contribute to XCI variation in aging, cancer, and development.

Overall, the range of X-Ra values was large (measured in percentage units, it had mean value of 75% in thyroid and 50% in breast cancer). Strong significant associations were observed across the study to the order of 1% unit increases in the biomarker, demonstrating its high-resolution power. Using longitudinal data, we also showed the high statistical reliability of X-Ra in tumor biopsies and blood, increasing its clinical applicability and translation to the clinic.

Our study has some limitations that need to be further addressed to improve the measurement X-Ra. For instance, X-Ra’s definition does not consider the individual likelihood of each XCI CpG to present 2-allele demethylation. In addition, our analyses do not account for the origin of the inactive X-chromosome, and therefore, the effect of skew inactivation and its relation to X-Ra should be considered in future studies.

Conclusions

Our work provides new insights into the risks and mechanisms associated with the deregulation of XCI in female cancer. We discovered that the level of XCI control varies in strength among women and decreases in cancer biopsies. Defective XCI is strongly associated with triple-negative breast cancer and worse survival. Defective XCI was reduced by chemotherapy treatment, suggesting potential new targets of intervention. The variability of XCI is also present in blood with validated correlations with cancer diagnosis. In a similar manner as mLOY is an important risk factor for cancer, several aging-related diseases and overall mortality in males, we show here that defective XCI is an important female-specific gonosome biomarker with long-range health consequences that are yet to be fully determined. Our study highlights the importance of considering inter-individual variation in XCI when investigating cancer risk and aging in females.

Materials and methods

Ra definition

DNA methylation is an important epigenetic modification that contributes to gene silencing on Xi45. Previous work has shown that the CpGs in the island-containing promoters of genes under XCI are methylated in Xi. Consequently, the mean methylation levels of these CpGs correspond to either 1-allele methylation values (methylation in Xi and no methylation in Xa) or 2-allele methylation values (methylation in both Xi and Xa)3,17. Therefore, a strong regulation of XCI should imply that no CpG in inactivated genes can have 2-allele demethylation values (hypomethylation in both Xi and Xa). After a thorough validation of this observation using multiple sources of data, we, consequently, propose to measure the level of defective Xi inactivation in a female biological sample by examining the number of these CpGs that are observed with 2-allele demethylation, i.e., methylation beta values < 0.2. Our hypothesis is that a consequence of a potential weakening of XCI is to allow the infrequent removal of methylation from any of the CpGs in the inactivated region of Xi. To differentiate between women with different quantitative levels of XCI, we compute the percentage of XCI CpGs that have 2-allele demethylated and should not have been under strong XCI. This percentage is then a measure of a patient’s defective XCI and/or a woman’s level of chromosome X reactivation (X-Ra), meaning the quantitative level of partial activation of Xi. Conceptually, the measure is a state variable that describes the level of XCI at a particular point in time and is given by the estimation of a woman’s probability that a CpG under XCI selected at random is observed with 2-allele demethylation

$${XRa}=P(M=0)$$

where M is the methylation state of a randomly selected CpG in Xi (0: demethylated, 1: methylated). Therefore, under strong XCI, X-Ra is expected to be low, and when XCI is weaken, X-Ra should increase.

We selected CpGs that map to X-linked genes with strong previously published evidence of consistent inactivation4,5. To allow wide applicability of X-Ra in different platforms, we used harmonized CpGs mapped in the TCGA workflow that included Human Methylation 27 (HM27), HumanMethylation 450 (HM450), and EPIC platforms. We thus created a final list of 4862 CpGs expected to have beta values in females higher than 0.2 due to XCI. The hypomethylation threshold of 0.2 surely captured CpGs that are clearly hypomethylated in different platforms13,46. This list of 4862 CpGs was our reference panel to compute X-Ra across studies with different methylation arrays. For each study, we then selected the CpGs that were in the reference panel and extracted the beta methylation levels. Then, for each individual, we defined the percentage of selected CpGs that had beta values lower than 0.2 as a measure of X-Ra.

TCGA data

The Cancer Genome Atlas (TCGA) consortium has mapped DNA methylation in thousands of cancer samples using Illumina Infinium Human Methylation 450 K BeadChip (Illumina 450 K array). We obtained methylation data for 15 different cancer sites and normal tissue. We downloaded data from ExperimentHub version 2.0.1 using the R package curatedTCGAData version 1.16.0. The cancer sites included breast (BRCA), bladder (BLCA), colon (COAD), head and neck (HNSC), kidney (KIRC), liver (LIHC), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), pancreas (PAAD), rectum (READ), thyroid (THCA), uterus (UCEC), skin (SKCM) and ovary (OV). We discarded the last two because they did not have methylation data on more than 5 samples of undiseased tissue. We analyzed a total of 11,232 CpGs on chromosome X for 3741 samples.

We downloaded clinical, tumor purity, RNA-seq and CNV data for BRCA to perform associations with X-Ra. For clinical associations, we downloaded data using TCGAbiolinks version 2.22.4 and fitted proportional hazards regression models for survival data adjusting by age for each independent cancer site. We also tested the association between X-Ra and hormonal status (ER, PR, HER2) of breast cancer, adjusting by age. For RNA-seq data, we inferred surrogate variables with svaseq from the SVA R library for each individual cancer. We tested the association of transcription levels of 20,532 genes with X-Ra adjusting by age and surrogate variables, transforming the expression levels with voom and fitting regression models with limma. In each cancer site, genes with less than 15 counts in more than 75% of the samples were removed from the analysis. Finally, we downloaded data of somatic mutations in BRCA and tested the association of X-Ra with the presence of recurrent mutations in tumors across genes.

TrueDiagnostic DNA biobank

The TruDiagnostic DNA biobank is a USA population-based cohort aged between 13 and 97 years old. The biobank includes 3590 individuals recruited between October 2020 and February 2022, who have chosen TruDiagnostic for DNA methylation analysis for biological aging. TruDiagnostic is an aging study in which a health questionnaire was administer at the time of the sample collection. Participants were screened over 120 clinical outcomes to determine their general health state. In particular, participants reported whether they had any cancer diagnosis, but no further details on treatment are available. Clinical data was collected, including anthropometric measures and diagnosis of common diseases. Telomere length data were available together with immune cell contamination. We analyzed methylation data from 1414 females of the 3259 individuals who passed quality control.

For the DNA methylation analysis, DNA was extracted from peripheral whole blood. The Infinium HumanMethylationEPIC BeadChip was used for DNA methylation assessment following the manufacturer’s protocol. Several quality controls and functional normalizations were performed using the meffil package, resulting in 745,150 probes, as described in detail elsewhere47. CpG sites were annotated to genes using the EPIC Illumina annotation ilm10b4.hg19. Blood cell types were estimated using the blood gse35069 reference panel from the meffil package. We analyzed a total of 4471 CpG sites from chromosome X. We assessed the association of X-Ra with age adjusting for leukocyte content and slide.

MESA study

We downloaded data from the Gene Expression Omnibus (GEO) (https://www.ncbi.nlm.nih.gov/geo/) with the reference GSE56046, to test the association between X-Ra and gene expression in blood, and validate the association with aging. This is a large population study with gene expression (Illumina HumanHT-12 v4 Expression) and methylation (Illumina HumanMethylation450) data in purified human monocytes from blood. Exclusion criteria included any serious medical condition which would prevent long-term participation. The data included covariates such as ancestry, gender, study site, and sample contamination with B-cells, T-cells, natural killer cells, and neutrophils. As ancestry, gender, and study site were encoded in a categorical variable with 18 different categories, we inferred sex from the mean expression of chromosome Y divided by the mean expression from the whole genome (Ry). Individuals from categories with Ry values less than −0.35 were classified as females. The classification was validated by the average methylation from inactive genes from chromosome X, where categories with values less than −0.5 were classified as females, thus completely matching those classified by Ry. We then analyzed data for 605 females with a mean age of 60 years (SD = 9.47). The methylation data were obtained in M values and transformed into Beta values [26]. We analyzed a total of 11,648 CpG sites from chromosome X. We assessed the association between X-Ra and age adjusting by leukocyte content and well. For gene expression data, we analyzed 47,308 transcripts and detected surrogate variables with SVA, adjusting by age, cohort and contamination with B-cells, T-cells, natural killer cells, and neutrophils.

Validation studies

We downloaded relevant methylation data from GEO to validate our findings in independent studies. We used methylation data of different cancer studies to validate X-Ra associations with cancer survival in TNBC (GSE78754, n = 70) and TNBC status (GSE225845, n = 224). We also analyzed studies of cancer recurrence for TNBC (GSE141441, n = 95) and a longitudinal study of chemotherapy intervention on 22 TNBC patients (GSE184159, n = 63).

We downloaded data from with reference GSE210255. The Genetic Epidemiology Network of Arteriopathy (GENOA) study is a large community-based study of hypertensive siblings, including African Americans from Jackson, USA, with methylation data obtained with the Illumina Infinium HumanMethylationEPIC BeadChip. After quality control and filtering, a total of 1394 samples were included in the analysis. We focused on female samples (n = 986) with a mean age of 57.2 years (SD = 10.3). We analyzed a total of 10,125 CpG sites on chromosome X and assessed the association between X-Ra and age, adjusting for plate and immune cell contamination and including a random effect on pedigree. Exclusion criteria included active malignancy.

We downloaded data from GSE142536 to validate the association between X-Ra in blood and cancer diagnosis. In this study, we also computed the test-retest reliability of the biomarker. The study included 17 elderly women, assayed immediately before major surgery, in the morning of postoperative day and between 4 and 7 days after surgery. Methylation data in blood was collected at each visit. Patients had a mean age of 79 years and three of them had cancer.

Statistical analyses

Data were analyzed using R version 4.1.1 and Bioconductor version 3.13. The goal was to determine the biological correlates of chromosome X reactivation (X-Ra) in studies with different characteristics and to establish the biological plausibility of the biomarker as a measure of defective X inactivation.

In the TCGA, for each of the 12 cancer sites, we inferred the value of X-Ra in constitutional and tumor tissues. We tested the association with cancer status (with logistic regression models) and survival (with proportional hazard regression models) across cancer sites, adjusting by age. We meta-analyzed the results with fixed and random effects models, using the rmeta library. The prediction power of X-Ra on cancer status was tested with a ROC area under the curve statistic. The optimal X-Ra was defined as the Youden distance, which was used to compute the attributable fraction of X-Ra of affected tissue on cancer risk, using the AFglm function of the AF library. For transcriptome-wide associations of X-Ra in each cancer site, we used limma and adjusted for SVA and age. Log2FC with their standard error were then meta-analyzed across cancer sites, and the associations were adjusted for multiple comparisons for the number of transcripts tested, taking Bonferroni significant threshold at 3.79 × 10−6. On these associations, enrichment analyses on biological processes were performed on genes with transcriptome-wide significance, using the enrichGO from clusterProfiler.

In the TCGA, we tested the association of X-Ra in tumor and healthy tissues with breast cancer risk and survival. For breast cancer risk, we fitted linear regression models of X-Ra on cancer status of the samples and adjusted for age. We also tested the association between X-Ra and ER, HER, and PER and triple-negative status, adjusting for age. Survival associations were performed with proportional hazard regression models adjusting for age. We also studied genome-wide associations of X-Ra with the frequency of copy number gains and deletions of breast cancer tissues in sliding windows of 0.5 Mb. To deal with windows with low frequency in copy number changes (1% at least), we fitted Bayesian regression models with bayesglm of arm, adjusting for age. Associations were adjusted for multiple comparisons using Bonferroni threshold of significance at P < 8.5 × 10−6. To determine associations with loss of function mutations (LOF), we downloaded a detailed catalog of somatic mutation data for TCGA from xenabrowser.net, and classified LOF mutations labeled as coding sequence variants, incomplete terminal codon variants, splice acceptor/donor variants, start lost, and stop gained or lost.

We also inferred X-Ra in blood and tested its association with age and markers of biological aging such telomere length, and Horvarth’s48 and Levine’s methylation age49, as computed by the R package methylclock50. Horvarth’s methylation clock uses 353 CpGs to estimate biological aging from methylation data, while Levine’s clock is based on 513 CpGs, none of which are in chromosome X. In the TruDiagnostic data, we tested the associations with age, markers of aging and with cancer diagnosis at any site. We fitted linear regression models, adjusting for immune cell abundance.

In the MESA study, we assessed the association between X-Ra and differential expression in purified monocytes from peripheral blood and with the age. In this study, the associations of X-Ra with transcriptomic signatures specific to XCI were also tested. The enrichment in associations (P < 0.001) of inactivated genes was tested using a hypergeometric (Fisher exact) test. For this study, the association between X-Ra and age (>65) was also tested using a linear regression model, adjusting for cohort and immune cell contamination.

Associations of X-Ra in blood with age were performed for the GENOA study (GSE210255). In GENOA, a linear model was fitted adjusting for plate and immune cell contamination. The regression coefficients for age across the three studies (MESA, GENOA, TruDiagnostic) were meta-analyzed using fixed and random effects models with two-sided tests using metagen from meta. Dichotomization of age was taken as the median age of cancer at any site according to SEER 22 2015–2019.

In the longitudinal study GSE142536 that included follow-up of individuals with cancer and controls, we fitted mixed models on X-Ra with condition, visit and age as fixed effects and subject as a random effect. We used the lmer R library. We also tested similar mixed models for the Levine’s methylation age, to contrast the effect of the condition between X-Ra and global methylation age. We also tested the significance of the intraclass correlation, to determine the test-retest reliability of X-Ra.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.