A comprehensive Xist interactome reveals cohesin repulsion and an RNA-directed chromosome conformation

iDRiP identifies multiple classes of Xist-interacting proteins

A systematic identification of interacting factors has been challenging because of Xist's large size, the expected complexity of the interactome, and the persistent problem of high background with existing biochemical approaches (20). A high background could be particularly problematic for chemical crosslinkers that create extensive covalent networks of proteins, which could in turn mask specific and direct interactions. We therefore developed iDRiP (identification of direct RNA interacting proteins) using the zero-length crosslinker, UV light, to implement an unbiased screen of directly interacting proteins in female mouse fibroblasts expressing physiological levels of Xist RNA (Fig. 1A). We performed in vivo UV crosslinking, prepared nuclei, and solubilized chromatin by DNase I digestion. Xist-specific complexes were captured using 9 complementary oligonucleotide probes spaced across the 17-kb RNA, with each probe being 25-nt length to maximize RNA capture while reducing non-specific hybridization. The complexes were then washed under denaturing conditions to eliminate factors that were not covalently linked by UV to Xist RNA. To minimize background due to DNA-bound proteins, a key step was inclusion of DNase I treatment before elution of complexes (See Supplemental Discussion). We observed significant enrichment of Xist RNA over highly abundant cytoplasmic and nuclear RNAs (U6, Jpx, 18S rRNA) in eluates of female fibroblasts (Fig. 1B). Enrichment was not observed in male eluates or with luciferase capture probes. Eluted proteins were subjected to quantitative mass spectrometry (MS), with spectral counting (21) and multiplexed quantitative proteomics (22) yielding similar enrichment sets (Table S1).

From three independent replicates, iDRiP-MS revealed a large Xist protein interactome (Fig. 1C; Table S1). Recovery of known Xist interactors, PRC2, ATRX, and HNRPU, provided a first validation of the iDRiP technique. Also recovered were macrohistone H2A (mH2A), RING1 (PRC1), and the condensin component, SmcHD1 — proteins known to be enriched on the Xi (23, 24, 19), but not previously shown to interact directly. More than 80 proteins were found to be ≥3-fold enriched over background; more than 200 proteins were ≥2-fold enriched (Table S1). Called on the basis of enrichment values and follow-up validations, novel high-confidence interactors fell into several functional categories: (i) Cohesin complex proteins, SMC1a, SMC3, RAD21, KIF4, PDS5a/b, as well as CTCF (25), which are collectively implicated in chromosome looping (26–28); (ii) histone modifiers such as aurora kinase B (AURKB), a serine/threonine kinase that phosphorylates histone H3 (29); RING1, the catalytic subunit of Polycomb repressive complex 1 (PRC1) for H2A-K119 ubiquitylation (23); and MINT/SPEN/RBM15/SHARP, which associates with HDACs; (iii) SWI/SNF chromatin remodeling factors; (iv) topoisomerases, TOP2a, TOP2b, and TOP1, that relieve torsional stress during transcription and DNA replication; (v) miscellaneous transcriptional regulators, MYEF2 and ELAV1; (vi) nucleoskeletal proteins that anchor chromosomes to the nuclear envelope, SUN2, Lamin-B receptor (LBR), and LAP2; (vii) nuclear matrix proteins, hnRPU/SAF-A, hnRPK, and MATRIN3; and (viii) the DNA methyltransferase, DNMT1, known as a maintenance methylase for CpG dinucleotides (30). In many cases (e.g., cohesins, SWI/SNF), multiple subunits of the epigenetic complex were identified, boosting our confidence in them as interactors. Because of stringent denaturing washes, non-crosslinked subunits of identified complexes were also of interest. We verified the interactions by performing a test of reciprocity: By using candidate proteins as bait in a RIP-qPCR of UV-crosslinked cells, we similarly detected the RNA-protein interaction (Fig. 1D).

To study their function, we first performed RNA immunoFISH of female cells and observed several patterns of Xi coverage relative to the surrounding nucleoplasm (Fig. 1E). Like PRC2, RING1 (PRC1) has been shown to be enriched on the Xi (23) and is therefore not pursued further. TOP1 and TOP2a/b appeared neither enriched nor depleted on the Xi (100%, n>50 nuclei). AURKB showed two patterns of localization — peri-centric enrichment (20%, n>50) and a more diffuse localization pattern (80%, data now shown), consistent with its cell-cycle dependent chromosomal localization (29). On the other hand, while SUN2 was depleted on the Xi (100%, n=52), it often appeared as pinpoints around the Xi in both day 7 differentiating female ES cells (establishment phase; 44%, n=307) and in fibroblasts (maintenance phase; 38.5%, n=52), consistent with SUN2's function in tethering telomeres to the nuclear envelope. Finally, the cohesins and SWI/SNF remodelers showed an unexpected “hole” or depletion relative to the surrounding nucleoplasm (100%, n=50–100). These patterns suggest that the Xist interactors operate in different XCI pathways.

To ask if the factors intersect the PRC2 pathway, we stably knocked down (KD) top candidates using shRNAs (Table S2) and performed RNA immunoFISH to examine trimethylation of histone H3-lysine 27 (H3K27me3; Fig. 2A,B). No major changes to Xist localization or H3K27me3 were evident in d7 ES cells (Fig. S1). There were, however, long-term effects in fibroblasts: The decreased in H3K27me3 enrichment in shSMARCC1 and shSMARCA5 cells (Fig. 2A,B) indicated that SWI/SNF interaction with Xist is required for proper maintenance of PRC2 function on the Xi. Steady state Xist levels did not change by more than 2-fold (Fig. 2C) and were therefore unlikely to be the cause of the Polycomb defect. Knockdowns of other factors (cohesins, topoisomerases, SUN2, AURKB) had no obvious effects on Xist localization and H3K27me3. Thus, whereas the SWI/SNF factors intersect the PRC2 pathway, other interactors do not overtly impact PRC2.

Xi-reactivation via targeted inhibition of three interactors

Given the large number of interactors, we created a screen to analyze effects on Xi gene expression. We derived clonal fibroblast lines harboring a transgenic GFP reporter on the Xi (Fig. S2) and shRNAs against Xist interactors. Knockdown of any one interactor did not reactivate GFP by more than 4-fold (Fig. 3A, shControl+none; Fig. S3A). Suspecting synergistic repression, we targeted multiple pathways using a combination drugs. To target DNMT1, we employed the small molecule, 5'-azacytidine (aza)(30) at a nontoxic concentration of 0.3 μM (≤IC₅₀) which minimally reactivated GFP (Fig. 3A, shControl + aza). To target TOP2a/b (31), we employed etoposide (eto) at 0.3 μM (≤IC₅₀), which also minimally reactivated GFP (Fig. 3A, shControl + eto). Combining 0.3 μM aza + eto led to an 80- to 90-fold reactivation — a level that was almost half of GFP levels on the Xa (Xa-GFP, Fig. 3A), suggesting strong synergy between DNMT1 and TOP2 inhibitors. Using aza + eto as priming agents, we designed triple-drug combinations inclusive of shRNAs for proteins that have no specific small molecule inhibitors. In various shRNA + aza + eto combinations, we achieved up to 230-fold GFP reactivation — levels that equaled or exceeded Xa-GFP levels (Fig. 3A). Greatest effects were observed for combinations using shSMARCC2 (227x), shSMARCA4 (180x), and shRAD21 (211x). shTOP1 and shCTCF were also effective (175x, 154x). Combinations involving remaining interactors yielded 63x to 94x reactivation.

We then performed allele-specific RNA-seq to investigate native Xi genes. In an F1 hybrid fibroblast line in which the Xi is of Mus musculus (mus) origin and the Xa of Mus casteneus (cas) origin, >600,000 X-linked sequence polymorphisms enabled allele-specific calls (32). Two biological replicates of each of the most promising triple-drug treatments showed good correlation (Fig. S4–S6). RNA-seq analysis showed reactivation of 75–100 Xi-specific genes in one replicate (Fig. 3B) and up to 200 in a second replicate (Fig. S3B), representing a large fraction of expressed X-linked genes, considering that only ~210 X-linked genes have an FPKM≥1.0 in this hybrid fibroblast line. Heatmap analysis demonstrated that, for individual Xi genes, reactivation levels ranged from 2x–80x for various combinatorial treatments (Fig. 3C). There was a net increase in expression level (ΔFPKM) from the Xi in the triple-drug treated samples relative to the shControl+aza+eto, whereas the Xa and autosomes showed no obvious net increase, thereby suggesting direct effects on the Xi as a result of disrupting the Xist interactome. Reactivation was not specific to any one Xi region (Fig. 3D). Most effective were shRAD21, shSMC3, shSMC1a, shSMARCA4, shTOP2a, and shAURKB drug combinations. Genic examination confirmed increased representation of mus-specific tags (red) relative to the shControl (Fig. 3E). Such allelic effects were not observed at imprinted loci and other autosomal genes (Fig. S7), further suggesting Xi-specific allelic effects. The set of reactivated genes varied among drug treatments, though some genes (Rbbp7, G6pdx, Fmr1, etc.) appeared more prone to reactivation. Thus, the Xi is maintained by multiple synergistic pathways and Xi genes can be reactivated by targeted inhibition of three distinct Xist interactors.

Xist interaction leads to cohesin repulsion

To investigate mechanism, we focused on one group of interactors — the cohesins — because they are implicated in establishing chromosomal architecture and their knockdowns destabilized Xi repression (Fig. 3). To obtain Xa and Xi binding patterns, we performed allele-specific ChIP-seq for two cohesin subunits, SMC1a and RAD21, and for CTCF, which works together with cohesins (33, 34, 28, 35). In wildtype cells, CTCF binding was enriched on Xa (cas), but also showed a number of Xi (mus)-specific sites (Fig. 4A)(36, 25). Allelic ratios ranged from equal to nearly complete Xa or Xi skewing (Fig. 4A). For the cohesins, 1490 SMC1a and 871 RAD21 binding sites were mapped onto ChrX in total, of which allelic calls could be made on ~50% of sites (Fig. 4B,C). While the Xa and Xi each showed significant cohesin binding, Xa-specific greatly outnumbered Xi-specific sites. For SMC1a, 717 sites were called on Xa, of which 589 were Xa-specific; 203 sites were called on Xi, of which 20 were Xi-specific. For RAD21, 476 sites were called on Xa, of which 336 were Xa-specific; 162 sites were called on Xi, of which 18 were Xi-specific. Biological replicates showed similar trends (Fig. S8A,B).

Cohesin's Xa preference was unexpected in light of Xist's physical interaction with cohesins, which suggested that Xist might attract cohesins to the Xi. We therefore conditionally ablated Xist from the Xi (Xi^ΔXist) and repeated ChIP-seq analysis in the Xi^ΔXist/Xa^WT fibroblasts (37). Surprisingly, Xi^ΔXist acquired 106 SMC1a and 48 RAD21 sites in cis, at positions that were previously Xa-specific (Fig. 4C,D). Biological replicates trended similarly (Fig. S8–S9). In nearly all cases, acquired sites represented a restoration of Xa sites, rather than binding to random positions. By contrast, sites that were previously Xi-specific remained intact (Fig. 4C,E, S8B), suggesting that they do not require Xist for either recruitment or maintenance. The changes in cohesin peak densities were Xi-specific and significant (Fig. 4F). Cohesin restoration occurred throughout Xi^ΔXist, resulting in domains of biallelic binding (Fig. 4G,S10–S12), and often favored regions that harbor genes that escape XCI (e.g., Bgn)(38, 39). There were also shifts in CTCF binding, more noticeable at a locus-specific level than at a chromosomal level (Fig. 4A,G), suggesting that CTCF and cohesins do not necessarily track together on the Xi. The observed dynamics were ChrX-specific and were not observed on autosomes (Fig. S13). To determine whether there were restoration hotspots, we plotted restored SMC1a and RAD21 sites (Fig. 4H; purple) on Xi^ΔXist and observed clustering within gene-rich regions. We conclude that Xist does not recruit cohesins to the Xi-specific sites. Instead, Xist actively repels cohesins in cis to prevent establishment of the Xa pattern.

Xist RNA directs an Xi-specific chromosome conformation

Cohesins and CTCF have been shown to facilitate formation of large chromosomal domains called TADs (topologically associated domains)(27, 40, 34, 28, 35, 41, 42). The function of TADs is currently not understood, as TADs are largely invariant across development. However, X-linked TADs are exceptions to this rule and are therefore compelling models to study their function (43–46). By carrying out allele-specific HiC-seq, we asked whether cohesin restoration altered the chromosomal architecture of Xi^ΔXist. First, we observed that, in wildtype cells, our TADs called on autosomal contact maps at 40-kb resolution resembled published composite (non-allelic) maps (27)(Fig. 5A, bottom). Our ChrX contact maps were also consistent, with TADs being less distinct due to a summation of Xa and Xi reads in the non-allelic (composite) profiles (Fig. 5A, top). Using the 44% of reads with allelic information, our allelic analysis yielded high-quality contact maps at 100-kb resolution by combining replicates (Fig. 5B) or at 200-kb resolution with a single replicate (Fig. S14). In wildtype cells, we deduced 112 TADs at 40-kb resolution on ChrX using the method of Dixon et al.(27). We attempted TAD calling for the Xi on the 100 kb contact map, but were unable to obtain obvious TAD, suggesting the 112 TADs are present only on the Xa. Thus, while the Xa is topologically organized into structured domains, the Xi is devoid of such organization across its full length.

When Xist was ablated, however, TADs were restored in cis and the Xi reverted to an Xa-like conformation (Fig. 5B, S14). In mutant cells, ~30 TADs were gained on Xi^ΔXist in each biological replicate. Where TADs were restored, Xi^ΔXist patterns (red) became nearly identical to those of the Xa (blue), with similar interaction frequencies. These Xi^ΔXist regions now bore little resemblance to the Xi of wildtype cells (Xi^WT, orange). Overall, the difference in the average interaction scores between Xi^WT and Xi^ΔXist was highly significant (Fig. 5C, S15A). Intersecting TADs with SMC1a sites on Xi^ΔXist revealed that 61 restored cohesin sites overlapped restored TADs (61 did not overlap). TADs overlapping restored peaks had larger increases in interaction scores relative to all other TADs (Fig. 5D,S15B) and we observed an excellent correlation between the restored cohesin sites and the restored TADs (Fig. 5E,S15C), consistent with a role of cohesins in re-establishing TADs following Xist deletion. Taken together, these data uncover a role for RNA in establishing topological domains of mammalian chromosomes and demonstrate that Xist must actively and continually repulse cohesins from the Xi, even during the maintenance phase, to prevent formation of an Xa chromosomal architecture.