JMJD3 intrinsically disordered region links the 3D-genome structure to TGFβ-dependent transcription activation

ARTICLE JMJD3 intrinsically disordered region links the 3D-genome structure to TGFβ-dependent transcription activation Marta Vicioso-Mantis1,7, Raquel Fueyo1,5,7, Claudia Navarro1, Sara Cruz-Molina 2, Wilfred F. J. van Ijcken 3, Elena Rebollo4, Álvaro Rada-Iglesias 2,6 & Marian A. Martínez-Balbás 1✉ Enhancers are key regulatory elements that govern gene expression programs in response to developmental signals. However, how multiple enhancers arrange in the 3D-space to control the activation of a specific promoter remains unclear. To address this question, we exploited our previously characterized TGFβ-response model, the neural stem cells, focusing on a ~374 kb locus where enhancers abound. Our 4C-seq experiments reveal that the TGFβ pathway drives the assembly of an enhancer-cluster and precise gene activation. We discover that the TGFβpathway coactivator JMJD3 is essential to maintain these structures. Using live-cell imaging techniques, we demonstrate that an intrinsically disordered region contained in JMJD3 is involved in the formation of phase-separated biomolecular condensates, which are found in the enhancer-cluster. Overall, in this work we uncover novel functions for the coactivator JMJD3, and we shed light on the relationships between the 3D-conformation of the chromatin and the TGFβ-driven response during mammalian neurogenesis. https://doi.org/10.1038/s41467-022-30614-y OPEN 1Department of Molecular Genomics, Instituto de Biología Molecular de Barcelona (IBMB), Consejo Superior de Investigaciones Científicas (CSIC), Barcelona 08028, Spain. 2Center for Molecular Medicine Cologne (CMMC), University of Cologne, Robert-Koch-Strasse 21, 50931 Cologne, Germany. 3Center for Biomics, Erasmus University Medical Center Rotterdam, Rotterdam, The Netherlands. 4Molecular Imaging Platform, Instituto de Biología Molecular de Barcelona (IBMB), Consejo Superior de Investigaciones Científicas (CSIC), Barcelona 08028, Spain. 5 Present address: Department of Chemical and Systems Biology, Stanford School of Medicine, Stanford University, Stanford, CA 94305, USA. 6 Present address: Institute of Biomedicine and Biotechnology of Cantabria (IBBTEC), CSIC/University of Cantabria, Santander, Spain. 7 These authors contributed equally: Marta Vicioso-Mantis, Raquel Fueyo. ✉email: [email protected] NATURE COMMUNICATIONS | (2022) 13:3263 | https://doi.org/10.1038/s41467-022-30614-y |www.nature.com/naturecommunications 1 1234567890():,; During mammalian neurogenesis, neural stem cell (NSC) progenitors differentiate into neurons in response to different signaling pathways1. Upon developmental pathway activation, transcription factors are recruited to the chromatin, and together with epigenetic regulators, they activate cisregulatory elements that will establish cell-specific gene expression patterns2–4. In mammals, promoters are usually regulated by more than one enhancer, and in fact, the number of enhancers in the mouse genome is one order of magnitude larger than the number of promoters5,6. This complex and sometimes-redundant configuration is crucial to ensure precise spatial-temporal control of the gene expression. However, how these multiple enhancers are orchestrated to regulate their target genes is still an open debate. Clusters of enhancers, also named super-enhancers by others7, are regions of euchromatin that are characterized by a high density of binding motifs, where transcription factors and cofactors such as Mediator, RNA-polymerase II (RNAPII) or chromatin remodelers colocalize (for review8). In recent times, it has been proposed that these clusters of enhancers facilitate transcriptional activation by promoting liquid-liquid phase separation (LLPS), a process by which molecules are condensed and concentrated in membrane-less compartments9–11. These condensates have been proposed to be formed by dynamic and weak multivalent interactions, that are characteristic of proteins that contain intrinsically disordered regions (IDR)12–18. Thus, intrinsically disordered proteins or regions have been suggested to drive the formation or to be incorporated into these biomolecular condensates. Despite the importance of enhancer clusters in cell identity establishment, we are still far from totally understanding the mechanisms by which they control gene transcription. Many research articles have revealed the importance of the 1D and 3D structure of the chromatin in development (reviewed in19–22). Nonetheless, the field lacks a specific assessment of the impact of individual developmental pathways on chromatin re-organization and function of specific loci. To fill this gap, we have analyzed the chromatin reorganization that underlies the transforming growth factor beta (TGFβ) pathway activation during neuronal commitment. We and others have demonstrated that in response to TGFβ, neural progenitors lose multipotency and commit to the neuronal lineage both in vivo and in vitro23–26. To do that, SMAD2/3, the major effectors of the pathway, cooperate with specific cofactors to regulate transcription. Particularly, SMAD3 interacts with the lysine demethylase (KDM) JMJD323,27,28, a Jumonji C (JmjC) domain-containing enzyme that catalyzes the histone 3 lysine 27 trimethylation (H3K27me3) removal29,30 and has been linked to numerous developmental processes (reviewed in31,32). In cortical progenitor cells, we have previously shown that JMJD3 cooperates with the TGFβpathway to induce neuronal differentiation23,33. In this context, JMJD3 and SMAD3 together bind and trigger the activation of neural cis-regulatory elements, presumably guided by the pioneer lineage-specific transcription factor ASCL1, and cooperating with the chromatin remodeler CHD833. Although some of the linear molecular components involved in the TGFβ-mediated enhancer activation have been identified, the relevance of their interactions at the 3D-level is still to be uncovered. Here, we perform 4C-seq experiments, and we illustrate that TGFβdrives enhancer-enhancer contacts that facilitate an enhancer cluster assembly, and ultimately gene activation. Upon TGFβstimulation, we observe that the establishment of multienhancer interactions requires the coactivator JMJD3. Using livecell imaging and molecular biology techniques, we demonstrate that a proline-rich IDR contained in JMJD3 is essential to induce LLPS, and we report a correlation between the JMJD3-containing molecular condensates and the enhancer driven gene activation. With our work, we reveal that JMJD3, containing a disordered domain, lies at the edge of chromatin structure and function upon TGFβstimulation of NSCs. Results TGFβdrives enhancer cluster assembly. The three-dimensional proximity between cis-regulatory regions has been systematically described as an intrinsic feature of the genome organization6. Nonetheless, the impact that the genome structure exerts over its function is still an open debate22. Within this framework, we hypothesize that the 3D-structure of the chromatin could be playing a role in the signal-dependent regulation of the TGFβresponsive enhancers. To test our hypothesis, we drew upon our well-characterized TGFβmodel of study, the E12.5 mouse NSCs. In these cells, the TGFβsignaling pathway is moderately active under basal conditions to allow neural progenitor proliferation23–26; however, further TGFβstimulation leads to the activation of hundreds of enhancers and genes that induce neuronal commitment in vitro and in vivo23–26,33 (Fig. 1a). Using this system, we asked whether the activation of the enhancers that occurs upon TGFβstimulation entails 3D-chromatin changes. For this purpose, we performed 4C-seq experiments using as a viewpoint (VP) a TGFβ-regulated enhancer that lies 38 kb downstream of the carbohydrate sulfotransferase 8 (Chst8) gene. The rationale to select this gene was the following: first, Chst8 is robustly upregulated upon TGFβtreatment; as indicated in Fig. 1b, the mRNA levels of the Chst8 gene increase up to ~25fold upon TGFβ-stimulation. Second, the distance between the VP and the Chst8 gene allows for a reliable resolution in 4C-seq experiments (38 kb) (Fig. 1c), as it is a known-fact that one of the 4C-seq caveats is the preferential ligation of the VP with its 1D closest regions, making the contacts that appear adjacent to the bait problematic to interpret. Third, Chst8 is a moderately long gene (138 kb), thus permitting the analysis not only of the contacts at the promoter level, but also potential interactions between the VP enhancer and the Chst8 gene body. Before analyzing which regions contact the selected VP, we confirmed that the VP is a TGFβ-responsive Chst8 enhancer. For that purpose, we used CRISPR/Cas9 technology to delete the VP enhancer (Fig. 1d) and measured the enhancer activity and the Chst8 transcriptional response to TGFβ-stimulation. To this end, we evaluated the transcription of enhancer RNA (eRNA), which serves as a readout of enhancer activation34. Results in Fig. 1e show a remarkable decrease of both enhancer activity and Chst8 induction upon TGFβ-stimulation in the Chst8 enhancer-deleted (ΔChst8 Enh) cells, demonstrating that the VP enhancer is an essential cisregulatory element of the Chst8 gene. After testing that our selected VP is a bona fide enhancer of the Chst8 gene, we performed two independent biological replicates of a 4C-seq experiment, where we analyzed the 3D-interactions between the VP and the genome before and after 3 h of TGFβ addition. The quality of the experiments was assessed following the criteria described in35 (Supplementary Data 1). The UCSC browser capture in Fig. 2a shows the obtained profiles for untreated and TGFβtreated NSCs. As expected, the proportion of cells displaying contacts between the VP enhancer and the Chst8 promoter increased upon TGFβtreatment (see region under light orange). However, we also observed novel contacts between the VP and the Chst8 gene body, these contacts were not particularly characterized by any type of regulatory element, but they suggest that TGFβtriggers a re-organization of the chromatin at this region (Fig. 2a and Supplementary Fig. 1a). Indeed, in the two biological replicates, the height of peaks located 500 kb upstream or downstream of the VP displayed a significant increase when NSCs cells were treated with TGFβ(p-value 0.0208), pointing to ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-30614-y 2NATURE COMMUNICATIONS | (2022) 13:3263 | https://doi.org/10.1038/s41467-022-30614-y | www.nature.com/naturecommunications TGFβas a driver of cis-regulatory region contacts (Fig. 2b and Supplementary Fig. 1b). Interestingly, in addition to the contacts observed between the VP and the Chst8 gene, we also identified several contacts occurring between the VP and interand intragenic enhancers located within the Pepd gene, a gene that lies 63 kb far from the VP, at its telomeric part (see regions under yellow in Fig. 2c). Surprisingly, the number and intensity of contacts between the VP and the enhancers located at its downstream region were higher than between the VP and the Chst8 gene promoter. The gene Pepd is not regulated by TGFβ(Fig. 1b), but its intragenic VP-contacting enhancers are bound by the TGFβtranscription factor SMAD3 upon TGFβtreatment (Fig. 2c), pointing to a structural role of Pepd in the convergence of TGFβ-regulated enhancers that could potentially be cooperating to activate TGFβresponsive gene promoters. These results indicate that TGFβ drives enhancer-enhancer contacts that lead to the assembly of an enhancer cluster; we named this assembly Chst8 enhancer cluster (EC). To confirm that the identified contacting regions within the Chst8 EC were TGFβ-responsive enhancers, we analyzed whether they became active upon TGFβpathway induction. To this end, we evaluated the transcription of eRNAs by qPCR upon TGFβ addition. We named the different enhancers of the EC enhancer (E)1, E2, E3, E4 and VP (Fig. 2c, d). Results in Fig. 2d show that the tested regions transcribed eRNAs in response to TGFβ.To prove the TGFβ-dependency of the EC activation we tested the eRNAs transcription in cells lacking the TGFβpathway effector SMAD3, that were previously characterized by our lab33. To this end, we measured the eRNA molecules transcribed from these enhancers upon TGFβtreatment in the control cells and in the SMAD3 depleted cells (shSMAD3). In concordance with the previous results, eRNAs were hardly induced in the shSMAD3 cells compared to the control cell line (Fig. 2d). Similarly, the Fig. 1 The VP is an essential Chst8 enhancer. a Schematic view of the model used in this study. NSCs were dissected from cerebral cortices of mouse fetal brains (E12.5) and cultured ex vivo (see methods). TGFβaddition leads to neuronal commitment. bNSCs were treated with TGFβ. Total RNA was prepared and the levels of the mRNA of the indicated genes were determined by qPCR. Values were normalized to the housekeeping gene Gapdh and the figure shows values relative to time 0 h. Results are the mean of three biologically independent experiments. Data are presented as mean values +/−SEM. **p< 0.01 (Pvalues were calculated using two-tailed Student’sttest, p=0.001131677 and p=0.006143072). Source data are provided as a Source Data file. cUCSC captures showing the chromatin landscape and SMAD3 binding around the Chst8 gene promoter and the Chst8 putative enhancer (VP) in NSCs. Tracks display ChIP-seq in NSCs treated with TGFβ(SMAD3) or untreated NSCs (H3K4me1, H3K27ac, and H3K4me3). Promoter and VP enhancer are shaded in light orange and yellow respectively. dSchematic representation of the CRISPR/Cas9 experimental approach used to delete the Chst8 putative enhancer in NSCs. Two gRNAs flanking the Chst8 enhancer region were used to create the deletion (2.9 kb). Red arrows represent primers to test the deletion. PCRs using Chst8 deletion and G6pd2 pairs of primers are shown at the bottom of the figure in parental and ΔChst8 enh NSC lines. Results are representative of three independent experiments. Source data are provided as a Source Data file. eParental and ΔChst8 enh cell lines were treated with TGFβfor 6 h. Total RNA was prepared and the levels of eRNA of the VP enhancer (left) or Chst8 mRNA (right) were determined by qPCR. mRNA and eRNA levels of Fapb4 were used as a control. Values were normalized to the Gapdh gene, and figure shows values relative to parental line. Data are presented as mean values +/−SEM. Results are representative of three biological independent experiments. ***p< 0.001 (Pvalues were calculated using two-tailed Student’sttest, p=2.2027E-05 and p=1.71676E-08). Source data are provided as a Source Data file. NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-30614-y ARTICLE NATURE COMMUNICATIONS | (2022) 13:3263 | https://doi.org/10.1038/s41467-022-30614-y| www.nature.com/naturecommunications 3 Chst8 gene was not expressed upon TGFβaddition in the shSMAD3 NSCs (Fig. 2e). Altogether, these results demonstrate that TGFβ-driven gene activation entails a reorganization of the chromatin structure. Moreover, this reorganization results in the formation of the Chst8 EC that is potentially involved in the regulation of genes upon TGFβ. TGFβ-mediated enhancer-cluster assembly depends on JMJD3. Previous work from our lab has demonstrated that the histone KDM JMJD3 functions as a cofactor for SMAD3 in the TGFβdriven activation of neuronal enhancers in NSCs33. For this reason, we decided to test whether JMJD3 was also occupying the enhancers involved in the Chst8 EC by analyzing our previously published JMJD3 ChIP-seq performed upon TGFβstimulation23. As shown in Fig. 3a, all the contacting regions belonging to the Chst8 EC (VP, E1, E2, E3 and E4) are occupied by JMJD3, consistent with the presence of enhancers at these regions. Next, motivated by the fact that the demethylase catalytic activity of JMJD3 is not involved in enhancer activation in our model33,we decided to address whether JMJD3 could be playing a structural role at enhancers, contributing to the Chst8 EC assembly. For this purpose, we efficiently depleted JMJD3 from NSCs (shJMJD3 NSCs) using lentivirus containing JMJD3-specific shRNAs (Supplementary Fig. 2a and refs. 23,33), and then, we performed a 4C-seq assay upon TGFβtreatment using the Chst8 VP enhancer. Figures 3a–c, and Supplementary Fig. 2b, c show the striking effect that the depletion of JMJD3 causes in the 3D-structure of Fig. 2 TGFβdrives enhancer-cluster assembly. a UCSC Genome Browser 4C-seq profiles generated in NSCs before and upon TGFβaddition are shown at the Chst8 promoter and gene body. The light orange box indicates enhancer-promoter contact. The yellow box indicates the VP enhancer (dark arrow). bBoxplot displaying the averaged values obtained from two biological independent replicates of RPM signals of the peaks located 500 kb around the VP - excluding the nearest ± 20 kb - (mm10 chr7:33841896-35860773) in NSCs untreated or treated for 3 h with TGFβ. An independent region located in another chromosome (mm10 chr4:33076383-35216108) was tested as a negative control. Boxes comprise values from Q1 to Q3 of the dataset; line corresponds to median value; whiskers show the data range (from min. to max. values within dataset). Depicted quantifications were performed for n =2 biologically independent samples. p-values are the result of a Wilcoxon-Mann–Whitney test. cUCSC Genome Browser captures showing 4C-seq profiles in NSCs untreated or treated (3 h) with TGFβspanning a 200 kb distance around the VP enhancer (dark arrow). ChIP-seq signals of SMAD3 are shown. The positions of enhancers (defined in33) are also displayed. The light orange box indicates enhancer-promoter contacts; yellow boxes show enhancerenhancer contacts. d,eThe top panel shows a scheme summarizing the enhancer-enhancer and enhancer-promoter contacts identified in the 4C-seq experiment. The bottom panel shows the treatment of control NSCs or shSMAD3 NSCs for 3 h with TGFβ.dshows eRNA levels from the indicated enhancers and eshows mRNA from the indicated genes quantified by RT-qPCR. Transcription values were normalized to the housekeeping gene Gapdh and the figure shows values relative to the untreated samples. Progesterone-responsive Fabp4 eRNA was used as a negative control. Results are the mean of three biologically independent experiments. Data are presented as mean values +/−SEM. *p< 0.05; **p< 0.01 (Pvalues were calculated using two-tailed Student’sttest, p=0.03229859 (E1), p=0.04082807 (VP), p=0.00839842 (E2), p=0.02618772 (E3), p=0.01592669 (E4) and p=0.00113168 (Chst8)). Source data are provided as a Source Data file. ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-30614-y 4NATURE COMMUNICATIONS | (2022) 13:3263 | https://doi.org/10.1038/s41467-022-30614-y | www.nature.com/naturecommunications the chromatin. Upon JMJD3 removal, we observed that the genomic contacts between the VP and the surrounding regions were severely reduced (Fig. 3c and Supplementary Fig. 2c). In particular, the contacts between enhancers belonging to the Chst8 EC region were abolished upon JMJD3 depletion (Fig. 3a, b); indicating that JMJD3 is directly or indirectly required for Chst8 EC assembly. Accordingly, Chst8, gene expression was markedly reduced (Fig. 3d). To broad our conclusions and to confirm that TGFβ-driven gene activation encompasses a chromatin structure reorganization that depends on JMJD3, we analyzed the 3D-chromatin status of other two candidate enhancers that potentially regulate the TGFβ-responsive and JMJD3-dependent genes, Ldlrad4 and Aopep23. These enhancers are located within the Ldlrad4 and Aopep genes, display SMAD3 and JMJD3 binding, and are surrounded by other SMAD3/JMJD3-bound enhancers, suggesting that they could potentially engage in enhancer clusters. Indeed, 4C-seq assays using these cis-regulatory regions as VPs (see quality of the experiments assessed as described in35, Supplementary Data 2 and 3) show contacts between the Ldlrad4 VP and the Ldlrad4 promoter (Supplementary Fig. 3a, shaded in orange), and between the VPs and the surrounding enhancers (Supplementary Fig. 3a, c, shaded in yellow). In agreement with our hypothesis, upon TGFβtreatment, the frequency of the contacts between cis-regulatory elements —enhancers or promoters—increased [Supplementary Fig. 3a, b (Ldlrad4), c, d (Aopep)]. Furthermore, these contacts remarkably diminished when JMJD3 was depleted (Supplementary Fig. 3a–d). Altogether these results corroborate that the TGFβpathway and JMJD3 are involved in 3D-chromatin structure regulation. As JMJD3 is a coactivator23,33,36, we decided to rule out the possibility of an indirect transcriptional effect triggered by the lack of JMJD3 in NSCs, that could potentially be affecting the expression of the proteins involved in loop formation37,38. To this end, we analyzed gene expression data from our previously published microarray experiments23, and we show in the Fig. 3 TGFβ-mediated enhancer-cluster assembly depends on JMJD3. a UCSC Genome Browser captures show 4C-seq profiles spanning 200 kb around the VP enhancer (black arrow) in NSCs untreated or treated (3 h) with TGFβ. ChIP-seq signals of SMAD3 and JMJD3 upon TGFβstimulation (0.5 and 3 h, respectively) are shown. The location of the members of the EC is also indicated with yellow boxes. bCapture showing a zoom into a region where TGFβinduced contacts are lost in JMJD3-depleted (shJMJD3) NSCs. cBoxplot displays the averaged values obtained from two biological independent replicates of RPM signals of the peaks located 500 kb around the VP - excluding the nearest ±20 kb - (mm10 chr7:33841896-35860773) in control or shJMJD3 NSCs untreated or treated with TGFβduring 3 h. An independent region located in another chromosome (mm10 chr4:33076383-35216108) was used as a negative control. n=2 biologically independent replicates were quantified. p-values are the result of a Wilcoxon-Mann–Whitney test. dControl NSCs or shJMJD3 NSCs were treated for 3 h with TGFβ. Then, total RNA was prepared, and the levels of the mRNA of the indicated genes were determined by qPCR. Values were normalized to the housekeeping gene Gapdh. The figure shows values relative to time 0 h. Results are the mean of three biologically independent experiments. Data are presented as mean values +/−SEM. **p< 0.01; **p< 0.001 (Pvalues were calculated using two-tailed Student’sttest, p=p=0.001131677 and p=0.003848794). Source data are provided as a Source Data file. NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-30614-y ARTICLE NATURE COMMUNICATIONS | (2022) 13:3263 | https://doi.org/10.1038/s41467-022-30614-y| www.nature.com/naturecommunications 5 Supplementary Fig. 4a that neither TGFβnor JMJD3 regulate the expression of some of the most characteristic proteins involved in loop formation (CTCF, SMC1/3, RAD21, PDS5A/B, and WAPL). Among these proteins, the Cohesin complex has been described to play a role in the establishment of dynamic contacts during gene transcription by being the main motor of the loop extrusion process39,40, thus, we analyzed the presence of the SMC1 subunit of the Cohesin complex in the Chst8 EC using previously published SMC1 ChIP-seq data in NSCs41. Interestingly, we noticed that some of the regions occupied by JMJD3 were also bound by SMC1 (Supplementary Fig. 4b). This observation prompted us to check whether JMJD3 co-occupies genomic regions with SMC1 in a genome-wide manner. The Venn diagram in Supplementary Fig. 4c shows that these proteins do not colocalize widely across the genome (only 19% of JMJD3 peaks overlap SMC1); this can be explained by the involvement of JMJD3 in transcription elongation, that leads to many JMJD3bound regions falling outside the cis-regulatory elements that engage in 3D-interactions36,42. Nonetheless, at the co-bound regions, JMJD3 and SMC1 demonstrate widespread peak overlapping (Supplementary Fig. 4d). The results described above indicate that JMJD3 is necessary for the establishment and/or maintenance of contacts between cis-regulatory regions. JMJD3 lacks DNA binding capacity, and structurally it only has one known domain, the demethylase catalytic domain JmjC. Previous work from our laboratory has shown that its demethylase domain is not required for the activation of a subset of enhancers in response to TGFβ33.To deeply understand whether demethylation of the H3K27me3 is involved in the Chst8 enhancer cluster activitation, we analyzed H3K27me3 ChIP-seq data from NSCs36. The results in Supplementary Fig. 5a show that the Chst8 locus lacks H3K27me3 prior to TGFβstimulation. Moreover, ChIP-qPCR experiments show that JMJD3 depletion did not lead to an increase in H3K27me3 levels neither at the Chst8 promoter nor at the scrutinized enhancers (Supplementary Fig. 5b). Of note, Supplementary Fig. 5b corroborates that the levels of H3K27me3 in these regions are negligible when compared to a classic H3K27me3-controlled promoter (Hoxd8), suggesting that changes on the H3K27me3 levels are unlikely a force driving the Chst8 EC formation or activation. JMJD3 is a highly disordered protein. As the investigated genomic loci are not marked by H3K27me3 prior to TGFβ activation, we hypothesized that JMJD3 could be impacting the 3D-structure of the chromatin through its unstructured domain. In the last years, numerous works have shed light on the impact on transcriptional regulation of these unstructured regions, named intrinsically disordered regions (IDR)43,44. With this in mind, we questioned whether JMJD3 with its unstructured domain could belong to the group of the intrinsically disordered proteins. To assess this, we analyzed the amino acid sequence of JMJD3 searching for disordered regions using the following previously validated algorithms: PONDR-VL345, IUPred46 and VSL247 (see methods). Overall, the three algorithms agreed on the highly significant disorder score of JMJD3 (Fig. 4a and Supplementary Fig. 6a, b). Specifically, PONDR-VL3 showed a median disorder score of 0.68 for JMJD3 (Fig. 4a), a value considerably higher than the 0.28 obtained when analyzing PSMA4, a wellknown structured protein used as an ordered protein control (Supplementary Fig. 6b). In addition, more than 70% of JMJD3 amino acids (71.39% using PONDR-VL3) exist in disordered domains (Fig. 4a and Supplementary Fig. 6a). Again, this value is higher than the proteasome component PSMA4, used as a negative control (23.37%) (Supplementary Fig. 6b). Looking at the different disordered fragments of JMJD3, we observed a remarkably long region with no defined structure that contains 943 amino acids (residues from 182 to 1125) and that was predicted to have a disorder score of 0.90, the highest observed in our data (Fig. 4a). From now on, we will refer to this region as JMJD3 IDR. In addition to disorder, the nature of the amino acid composition has also been shown to play an important role in IDR-mediated transcription regulation. Moreover, disordered regions frequently coincide with low-complexity domains that are biased for certain amino acids10,13,14,18. To check whether this was the case for JMJD3 we used the SEG algorithm48 (see methods) looking for JMJD3 complexity prediction. As shown in Fig. 4b, 30.4% of JMJD3 was predicted to contain low-complexity segments. Furthermore, by examining its amino acid composition, we found a remarkable abundance of prolines (24% of the total amino acids in the protein) (Fig. 4c and Supplementary Fig. 6c), that displays widespread conservation among mammals (Supplementary Fig. 6d). Strikingly, we found proline tracks as long as 20 residues (Supplementary Fig. 6c). Other structural features of the JMJD3 IDR are its high content of charged residues (21% of the protein) (Fig. 4c, d, and Supplementary Fig. 6c), its regions different to the proline tracks that also display high hydrophobicity (Fig. 4d), and its high serine content when compared to the average in the mouse proteome, and similar to other described IDRs (Fig. 4c)15,49. Proline residues have been described as highly hydrophobic amino acids whose concatenation generates sticky domains that bind rapidly and reversibly to other proteins50. It is known that hydrophobic interactions, as well as electrostatic ones, are relevant for biomolecular condensate formation. Thus, the amino acid composition of JMJD3 seemed to favor its potential to be involved in the socalled phase separation process9,51. Given this, we used catGRANULE52 and PSPredictor53 algorithms (see methods) to predict JMJD3 phase separation ability. Both tools returned high scores (0.83 and 0.99 respectively) for JMJD3 (Fig. 4e, f), similar to proteins known to be involved in phase separation (e.g. MED1, 0.99) and higher than the proteasome protein PSMA4 (0.001) (Fig. 4f). Altogether, these data point to JMJD3 as a highly disordered protein with the potential to undergo phase separation. JMJD3 undergoes LLPS in vitro and in vivo. LLPS is a physicochemical process that consists on the demixing of a fluid into a diluted phase and a dense phase. It is well known that proteins mediating phase separation contain IDRs, and it is starting to be uncovered the role that these IDRs play on transcription regulation as mediators of biomolecular condensation12–15,17,54–56. On the grounds of these recent discoveries, we hypothesized that JMJD3 could be contributing to the establishment of 3D-contacts by nucleating protein and nucleic acid scaffolds to form membrane-less condensates through LLPS. To test this idea, we decided to perform in vitro droplet assays using a construct that expressed JMJD3 fused to monomeric EGFP (mEGFP) and HA (Supplementary Fig. 7a). First, we tested that the resulting fluorescence protein had the predicted molecular weight and was well recognized by JMJD3 antibody when ectopically expressed (Supplementary Fig. 7a, b). Next, we expressed mEGFP-JMJD3 in HEK293T cells and performed an in vitro droplets assay using nuclear extracts. Our data shows that the mEGFP–JMJD3 protein forms droplets that do not appear when we overexpress mEGFP alone, reflecting that the droplets can be attributed to JMJD3, and not to the mEGFP tag (Fig. 5a). The droplets showed features [circularity, convexity and aspect ratio] that are characteristic of a liquid-like nature (Fig. 5a, bottom panels). Interestingly, some of ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-30614-y 6NATURE COMMUNICATIONS | (2022) 13:3263 | https://doi.org/10.1038/s41467-022-30614-y | www.nature.com/naturecommunications these in vitro droplets were shared with MED15, a well-known component of enhancers that forms nuclear condensates (Supplementary Fig. 7c, d). When we overexpressed mEGFP-JMJD3, mCherry-MED15 or both we observed that 57% of JMJD3 droplets colocalized with MED15 (Supplementary Fig. 7d). In fixed cells, overexpression of mEGFP–JMJD3 formed nuclear puncta (Fig. 5b), and the intensity of these puncta increased along with the amount of JMJD3 inside the cell (Fig. 5b); in fact, at 0.25 ug of plasmid overexpressed, we started to observe aggregates (see below). Next, we tested the sensitivity of these condensates to the aliphatic alcohol 1,6-hexanediol, a chemical compound that has been demonstrated to disrupt the hydrophobic interactions that sustain the phase-separated droplets57. We observed that the treatment led to a reduction in the number and size of JMJD3 puncta (Fig. 5c). Importantly, endogenous JMJD3 also formed nuclear puncta as detected by immunofluorescence using two different antibodies against JMJD3 in NSCs and HEK293T cells (Fig. 5b, without transfection, d and Supplementary Fig. 8a), ruling out the possibility that the observed puncta could be an overexpression artifact. These data suggest that JMJD3 condensates occur at endogenous levels and that these condensates represent a separated phase inside the cell. Liquid-like condensates have been suggested to exhibit a remarkable dynamic nature, and their internal molecules have been described to diffuse rapidly12,58. Based on this, we sought to analyze whether the JMJD3 puncta exhibited liquid-like properties by analyzing the rate of fluorescence recovery after photobleaching (FRAP) of the overexpressed mEGFPJMJD39,59. After photobleaching mEGFP-JMJD3 puncta recovered fluorescence almost completely on a time scale of seconds (Fig. 5e, Supplementary Fig. 8b and Supplementary Movie 1), in agreement with what is observed for other proteins that form either liquid-like condensates (BRD4 and MED1) or membraneless organelles60. We also calculated the mobile fraction (the molecular pool that undergo exchange within the FRAP zone), which corresponds to 0.95 for this protein. Moreover, the aggregates that appear when high levels of proteins are overexpressed (mentioned above) showed reduced mobility in FRAP assays (Supplementary Movie 2). These data suggest that JMJD3 droplets exhibit liquid-like properties and that its Fig. 4 JMJD3 is a highly disordered protein. a Disorder prediction of human JMJD3 using PONDR-VL3 algorithm. In the bottom panel, the disorder score and the lengths of the predicted disordered regions are indicated (length of disordered segments >50 amino acids). A schematic representation of JMJD3 described domains is shown on top of the graphic. bAnalysis of the presence of low-complexity domains in JMJD3 using the SEG algorithm. The percentage of low complexity regions is indicated on the right side. Low complexity regions are depicted in yellow. A schematic representation of JMJD3 described domains is shown on top of the panel. cAmino acid composition of JMJD3 IDR, JMJD3 catalytic domain, mouse proteome and disordered proteins defined by the presence of a 50 residues fragment whose IUPRED median score is at least 0.55 and that is not found in Pfam (so that functional domains are avoided). The percentages of acid, basic, hydrophobic, and hydrophilic amino acids of JMJD3 are shown on the right panel. dJMJD3 hydrophobicity profile was determined using the ExPASy website with the Hopp and Woods scale and a sliding window of 21. e,fThe potential of JMJD3 to phase separate was determined using catGRANULE (on the left) (e) and PSPredictor (on the right) (f) algorithms. NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-30614-y ARTICLE NATURE COMMUNICATIONS | (2022) 13:3263 | https://doi.org/10.1038/s41467-022-30614-y| www.nature.com/naturecommunications 7 conforming molecules exchange rapidly between the condensates and the surrounding. JMJD3 IDR is necessary for condensate formation. As the condensation of molecules into liquid-like droplets has been related to the IDRs present in the conforming proteins61,we chose to investigate whether the proline-rich IDR of JMJD3 is necessary for being part of biological condensates. To do this, we deleted the proline-rich IDR domain (amino acids 140–820) from our previously characterized mEGFP-JMJD3 plasmid, and we named this new construct mEGFP-JMJD3 ΔIDR (Fig. 6a). We ectopically expressed this protein (Fig. 6a) and analyzed its ability to form droplets in nuclear extracts. The results in Fig. 6b demonstrate that the mEGFP-JMJD3 ΔIDR protein did not form droplets in vitro. We next investigated the competence of mEGFP-JMJD3 ΔIDR to form puncta in fixed cells. These two versions of the protein were distributed between nucleus and cytoplasm, even though the JMJD3 ΔIDR shows some bias for the cytoplasm. Our immunofluorescence experiments revealed that the JMJD3 mutant was unable to form puncta (Fig. 6c). These data support that the IDR domain and probably the prolines are necessary for JMJD3 phase separation. Even though we acknowledge that identifying the precise amino acids involved in Fig. 5 JMJD3 undergoes LLPS in vitro and in vivo. a mEGFP and mEGFP–JMJD3 proteins were analyzed using droplet-formation assays in nuclear extracts at room temperature with 150 mM NaCl. Quantifications of the number of droplets per frame, circularity, convexity and aspect ratio (AR) are displayed. Data are the mean ± SEM. Boxes comprise values from Q1 to Q3 of the dataset; line corresponds to median value; whiskers show the data range (from min. to max. values within dataset). ***p< 0.001 (Student’sttest, p=1.09982E-06). Droplets in 5 fields in each group from three biologically independent experiments were quantified, n=150. Scale bar, 5 μm. bConfocal microscopy images of HEK293T cells transfected with mEGFP-JMJD3. Quantifications of the intensity of JMJD3 puncta are shown on the right. Data show the mean ± SEM. Boxes comprise values from Q1 to Q3 of the dataset; line corresponds to median value; whiskers show the data range. ***p< 0.001 (Student’sttest, p=9.98785E-06 and p=7.15724E-17). n=50 transfected cells in each group were quantified; Images are representative of 3 biologically independent experiments. Scale bar, 5 μm. Western blot displays the levels of overexpressed JMJD3. cHEK293T cells were transfected with 0.05 ug mEGFP–JMJD3, treated with 6% 1,6-HD for 5 min and imaged at 60 and 120 s. Nuclei were visualized with DAPI (blue). Quantification of the nuclear puncta per cell is shown on the right. Data are the mean ± SEM. *p< 0.05 (Student’sttest, p=0.0182428 and p=0.01162199). n=130 transfected cells were quantified; Images are representative of three biologically independent experiments. Scale bar, 5 μm. dNSCs and HEK293T cells were fixed,and endogenous JMJD3 was visualized by immunostaining assay. The images are representative of three biologically independent experiments. Scale bar, 5 μm. eFRAP assay in HEK293T cells expressing 0.05ug of mEGFP–JMJD3. Images are representative of three biological replicates. Quantification shows the curve fit results of FRAP data for mEGFP-JMJD3 to a double-exponential smoothing (R2=1), where bleaching events occurs at t=0 s. Data are plot as background-subtracted and normalized mean (n=27 cells). Scale bar, 5 μm. Source data are provided as a Source Data file. ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-30614-y 8NATURE COMMUNICATIONS | (2022) 13:3263 | https://doi.org/10.1038/s41467-022-30614-y | www.nature.com/naturecommunications the condensation would benefit our work, both the length and the complexity of the IDR domain prevented this analysis. Nonetheless, our data suggest that not only one type of amino acid but also several contribute to JMJD3 phase separation. Once known the relevance of the IDR domain, we analyzed if the catalytic activity of JMJD3 plays a role in the formation of condensates. To do that, we used a plasmid encoding JMJD3 mutated at the catalytic domain (JMJD3 HE > AA) (Supplementary Fig. 9a). This mutant lacks the capability of demethylating the H3K27me3 mark and functions as a dominant negative form of JMJD323. Experiments in fixed cells revealed that the JMJD3 HE > AA mutant formed nuclear puncta of the same intensity and volume than those of the wild type version of JMJD3, suggesting that the catalytic activity is not essential for condensate formation (Supplementary Fig. 9a, c). To further prove that the catalytic activity of JMJD3 is not required for condensate formation, we employed a specific inhibitor of the JMJD3 catalytic activity named GSK-J462. We treated the cells for 6 h, a period of time that was enough to effectively inhibit JMJD3 (Supplementary Fig. 9b). Using the GSK-J4 inhibitor, we observed puncta formation of the same intensity and volume than those of the non-treated cells (Supplementary Fig. 9b, c). Moreover, the Fig. 6 JMJD3 IDR is essential for condensate formation. a mEGFP–JMJD3 and mEGFP–JMJD3 ΔIDR expression vectors were transfected into HEK293T (0.05ug). 24 h later total protein extracts were prepared and the JMJD3 (HA) and TUBULIN levels were determined by immunoblot. The image shown is representative of two independent experiments. Source data are provided as a Source Data file. bmEGFP, mEGFP–JMJD3 and mEGFP-JMJD3 ΔIDR proteins were analyzed using droplet-formation assays in nuclear extracts at room temperature in the presence of 150 mM NaCl. Quantifications of the droplets are displayed on the right. Data are the mean ± SEM. Boxes comprise values from Q1 to Q3 of the dataset; line corresponds to median value; whiskers show the data range (from min. to max. values within dataset). ***p< 0.001 (Pvalues were calculated using one-tailed Student’sttest, p=0.000113231). Droplets in 5 fields in each group from three biologically independent experiments were quantified. Scale bar, 5 μm. Source data are provided as a Source Data file. cConfocal microscopy images of HEK293T cells transfected with 0.05 ug mEGFP-JMJD3 or mEGFP-JMJD3 ΔIDR. The images are representative of three biologically independent experiments. Quantifications of the number of JMJD3 puncta are shown on the right. Data show the mean ± SEM. Boxes comprise values from Q1 to Q3 of the dataset; line corresponds to median value; whiskers show the data range (from min. to max. values within dataset). ***p< 0.001 (Pvalues were calculated using one-tailed Student’sttest, p=2.5178E-09). n=20 transfected cells in each group were quantified. Scale bar, 5 μm. Source data are provided as a Source Data file. NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-30614-y ARTICLE NATURE COMMUNICATIONS | (2022) 13:3263 | https://doi.org/10.1038/s41467-022-30614-y| www.nature.com/naturecommunications 9