Identification of a discrete transcriptome for MCCs

The MCC is a major cell type in the Xenopus larval skin that forms in the early embryo along with mucus-secreting cells and proton-secreting ionocytes. The differentiation of these cell types can be readily analyzed using RNAseq[16] analysis on skin progenitors isolated away from the embryo. To identify gene expression associated with distinct differentiation programs, we analyzed isolated progenitors where cell fate was altered in a predictable way. Blocking Notch in these progenitors (by expressing a DNA binding mutant of suppressor of hairless, dbm[17], labeled here as “Notch-”) leads to a marked increase in the number of MCCs and ionocytes that arise within an inner layer and move into the outer epithelium (Fig 1B and 1C). Activating Notch, (by expressing the intracellular domain of Notch, icd[18], labeled here as “Notch+”) completely suppresses MCC and ionocyte formation. Ectopic expression of Multicilin[19] in skin progenitors along with active Notch completely rescues MCC differentiation, and moreover, converts most skin progenitors into MCCs, including those in the outer layer normally fated to become mucus-secreting cells. Inhibiting Multicilin activity using a dominant-negative mutation (termed dnMulticilin [19]) blocks the formation of MCCs, but spares other cell types such as ionocytes. Finally, for an additional comparison, we ectopically expressed Foxi1 in these progenitors, a transcription factor that promotes ionocyte differentiation alone. Epithelial progenitors were isolated from injected embryos at stage 10, and subjected to RNAseq analysis as the various differentiation programs unfolded (Fig 1C). To maximize the cell-specific signal in these data sets, we compared conditions in which the change in cell number was great: for example, progenitors with activated Notch (using Notch-icd) have virtually no MCCs, whereas progenitors co-injected with activated Notch and Multicilin have large numbers without a confounding ionocyte population. MCC and ionocyte number is also markedly increased by blocking Notch signaling (using dbm), whereas progenitors co-injected with dbm and dnMulticilin have virtually no MCCs but still have increased ionocytes. A complete comparison of the effects of these manipulations is shown in S1 Table. Clustering of genes with the greatest variance (see RNAseq informatics section in Methods) in these extensive RNAseq data sets revealed that the largest cluster of genes by far was the one that changed in association with MCC differentiation (Fig 1F; this group of conditions is detailed in S1 Table and shown in S1 Fig; the cluster of genes themselves is in S2 Table). By comparison, this cluster of genes associated with MCC differentiation dwarfed that associated with ionocyte differentiation (induced with Foxi1 as a control, Fig 1F, S1A Fig, S3 Table), with Notch signaling (Fig 1F, S1B Fig, S4 Table), or with embryonic age (Fig 1F, S1C and S1D Fig, S5 and S6 Tables). These data emphasize the relatively large number of genes coordinately upregulated during the formation of this specialized epithelial cell type.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1.

Identification of a MCC transcriptome (A) Confocal image of X. laevis skin showing a multiciliated cell (MCC), ionocytes (IC), and outer cells (green cells, unlabeled). (B) Differentiation of X. laevis skin. Multipotent progenitors are specified to become MCCs or ionocytes in the inner layer (red cells) by Notch signaling; they then intercalate into the layer of outer cells (green cells). (C) Diagrams illustrating how the numbers of MCCs and ICs in the skin change when Notch, Multicilin and/or Foxi1 activity is manipulated. (D) Schematic of the general experimental strategy used to analyze X. laevis epidermal progenitors (“cap”), after manipulating gene expression using RNA injection. (E) Venn diagram using multiple RNAseq experiments to define a core list of genes expressed in MCCs based on an intersectional strategy. (F) Heatmap of transcriptional variation across all experimental conditions and timepoints subjected to RNAseq analysis (see Methods for more details). For clarity of display, sample names are omitted here but can be seen in S1 Fig.

https://doi.org/10.1371/journal.pgen.1006538.g001

We next generated a reliable list of genes upregulated during MCC differentiation based on differential expression (p adjusted for multiple testing < 0.05) between pairs of conditions exhibiting the greatest change in MCC number (Fig 1B and 1E: dbm vs. Notch-icd, Notch-icd + Multicilin vs. Notch-icd, and dbm vs dbm + dnMulticilin). A conservative list of 950 genes was derived from the oldest, 9-hour timepoint (813 genes if we collapse the L and S forms based on orthology assignment); this list changed in all three comparisons (p<0.05, Fig 1E), and provides a high-confidence collection of the genes upregulated during MCC differentiation, which we designate “core MCC genes” (S7–S9 Tables). To assess this claim, we compared our core MCC gene list to those generated using other approaches, including Foxj1 overexpression in Zebrafish,[20] FACS sorting of Foxj1⁺ cells from mammalian airway epithelia,[21] or knockdown of Rfx2 in Xenopus.[22] These gene lists, as well as our own, were plotted using our data based on their level of expression (using normalized counts) versus the fold-change that occurred in our RNAseq analysis as MCC number increased. In each case, the published lists lacked a large fraction of genes on our list (red dots in S2B–S2D Fig, also see S11–S13 Tables for a comparison of overlap between these sets), and the genes unique to our list for each comparison were strongly enriched for cilia GO terms (S3A, S3C and S3E Fig). Conversely, the published list included a large number of genes that did not increase with more MCCs in our data, and thus absent from our list (grey dots in S2B–S2D Fig), but these genes were either not strongly enriched for any GO terms or enriched for more general terms (S3B, S3D and S3F Fig e.g., “biological process”). Altogether, these comparisons suggest that the core MCC list more accurately identifies genes upregulated during MCC differentiation, and thus serves as a solid foundation for the subsequent analysis described below.

3D chromosomal architecture suggests that MCC genes are enriched at TAD boundaries

We next related the MCC core genes to the 3-dimensional genome, in particular to areas of high self-interaction known as topologically-associated domains (TADs). To obtain 3D chromosomal interactions across the entire genome, we performed tethered conformation capture (TCC)[23] on epithelial progenitors isolated from X. laevis embryos, analyzing both wild-type tissue (unmanipulated progenitors containing a mixture of outer cells, ionocytes, and multiciliated cells), and after inducing MCC differentiation using Multicilin (resulting in progenitors consisting of MCCs). Most of the interactions in our data sets were within chromosomes (92%-87% of totals, respectively, S4 Fig), indicating a low level of interchromosomal interactions (high levels of these are thought to reflect spurious interactions[24]). The active and inactive compartments in the wild-type and Multicilin-injected samples were compared by computing Pearson’s correlation matrices[25] (Fig 2A). This analysis failed to detect loci with a negative correlation coefficient and very few with correlation coefficients below 0.5 (268 bins out of a possible 50,030, S14 Table), suggesting that while these two tissues are significantly different in their transcriptional profiles (Fig 1), there was very little in the way of large-scale changes, such as compartment switching between the two conditions.[7,25] Genes found in the few regions that were poorly correlated are not implicated in MCC differentiation (S14 Table), while those on our core MCC list, for example, wdr16, were highly correlated between the two conditions (Fig 2A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2.

TADs and their genomic features (A) Browser screenshot of the genomic region surrounding wdr16, a MCC expressed gene. Top track is correlation coefficient between 3D chromosome conformation data between wild-type ectoderm and ectoderm injected with Multicilin, middle tracks show ChIPseq results as labeled, bottom track is called topological domains. (B) Interaction matrix of tethered conformation capture of the same genomic region. High-throughput methods of determining 3-dimensional chromatin structure such as TCC or HiC involve isolating DNA-protein complexes either via dilution (classical HiC), by fixing the proteins to avidin beads (TCC) or in situ nuclear fixation (in situ HiC), cutting DNA in this folded state with a restriction enzyme at many positions, religating, and sequencing. Restriction sites near loops of DNA will ligate across the loops at some frequency, which can then be used to reconstruct the frequency of contact between two close or distant regions of the genome. Here, regions interacting across the genome more often than a linear model of DNA would predict are shown, with darker red indicating a higher frequency of interactions.[6–9] (C-F) Metagene plots showing the distribution of various features relative to all TADs. All domains are normalized to the same size, the domain region is in the center of each plot, and the two vertical lines denote the domain boundaries. Areas in outer edges of the plot denote flanking genomic regions of some 200 kb. Each quartile of the plot (one quartile upstream, two quartiles inside the domain region itself, and one quartile downstream) is broken into 175 bins, and each dot denotes the measured values of one bin.

https://doi.org/10.1371/journal.pgen.1006538.g002

At higher resolution, 3D interaction maps reveal TADs, structures thought to facilitate complex transcriptional regulation.[7,9,26,27] To determine the positions of TADs in the X. laevis genome, we pooled the reads from the two TCC experiments and calculated a directionality index,[26] resulting in 7,249 domains (Fig 2A and 2B). As TAD boundaries are reportedly enriched for CTCF and cohesin, we performed ChIPseq on the cohesin component Rad21[26,28] (Fig 2A, S5 Table) and found striking enrichment of this protein at our called domain boundaries, providing independent validation that they were called accurately (Fig 2A and 2C). We also found Rad21 peaks not associated with TADs (Fig 2A, ~81% of total peaks); reports by others suggest that both Rad21 and CTCF are frequently outside TADs and are most likely to mark TAD boundaries when cobound.[29] We annotated the positions of transcriptional start sites (TSS’s) relative to TADs and found enrichment at the boundaries (Fig 2C), in agreement with previous work on mammalian genomes.[26] We further performed ChIPseq on two histone modifications: H3K4me3, a mark associated with actively-transcribed promoters (Fig 2A, S6 Fig), and H3K27ac, a mark associated with active promoters and enhancers, including superenhancers[30] (Fig 2A, S7 Fig). The H3K4me3 ChIPseq (Fig 2A and 2D, S6 Fig) exhibited broadly similar nucleotide frequencies and peak shapes as other vertebrate promoters[31] and largely overlapped, within a 2 Kb window, our annotated 5’ ends of X. laevis gene models (16,231 out of 24,384 total models; note not all of these genes will be expressed in epithelial progenitors and would thus largely lack H3K4me3 peaks). When these features were overlaid onto the TAD map, the promoter regions, as defined by both annotated TSS’s and H3K4me3 peaks, tended to lie close to the TAD boundaries (Fig 2C and 2D), while enhancers were depleted at boundaries and instead enriched in the middle of the domains themselves (Fig 2D). We localized our core MCC genes onto the TAD map using H3K4me3 binding at their promoters (Fig 2E) or the positions of the core MCC TSS’s themselves (S8 Fig), finding that, like all genes, core MCC genes were enriched at TAD boundaries, rather than located within domains.

MCC genes reside in enhancer-sparse regions

The enrichment of MCC genes at TAD boundaries and far from the bulk of enhancers led us to examine how many enhancers were near these genes in greater detail. A single collection of all peaks with active histone marks (H3K4me3 and H3K27ac) was generated and assigned to genes with the closest TSS (S15 Table), based on the assumption that this approach estimates local enhancer density rather than attempting to predict targets of every enhancer. Using this approach, we found an average of 2.84 active peaks closer to a given gene than any other gene across the entire genome. Genes with the highest number of flanking active peaks (> 4) tended to be TFs based on GO term functional analysis (“transcription factor activity”, p value 1.02E-20 for genes with more than 4 flanking active peaks, S15 Table). To confirm this, we checked known TFs (Xenopus paralogs of 1,692 human TFs[32]) and found they had 4.04 active flanking peaks on average and were positioned inside TADs, where enhancer density was highest (S9 Fig). Conversely, genes with a single nearby active peak were enriched in GO terms such as “protein transport” (p value = 1.7E-6), and “ribosome”, (p value = 6.1E-5). Known housekeeping genes (Xenopus paralogs of 3,804 human genes[33]) contained 2.40 active flanking peaks on average, and these genes were also enriched at TAD boundaries rather than within them (S9 Fig). Finally, in a general survey, all genes with few enhancers largely resided at the boundaries, while genes with many enhancers were depleted at boundaries but instead located within domains themselves (Fig 2F). These results are consistent with the idea that housekeeping genes are located at TAD boundaries where enhancer density is low, while developmental regulated genes, such as TFs, are located in areas of high enhancer density within TADs. In this light, it is striking that the MCC-expressed genes averaged few peaks per gene (2.42) and were enriched at TAD boundaries (Fig 2E), similar to housekeeping genes. Thus, these data indicate that MCC genes, although expressed in a cell type-specific manner, resemble housekeeping genes, raising the question of how these genes become activated in this configuration.

TF binding motifs associated with MCC promoters

Previous studies have found that proximal regulatory regions of genes expressed in motile ciliated cells are enriched in Rfx binding sites[34,35] and that in Xenopus, Rfx2 binding is found extensively in association with cilia genes.[22] Consistent with these observations, the binding site for the Rfx proteins was by far the most enriched motif identified when sequences (+/-100 bases) around the TSS of the core MCC genes (verified by H3K4me3, S6 Fig) were examined for overrepresented motifs using an unbiased approach (Fig 3A).[3] We also detected a similar, striking enrichment of Rfx motifs in the promoters of MCC gene orthologs in human, mouse, zebrafish, fruit fly, and sea starlet (Fig 3B), despite the fact that the core promoter sequences vary substantially across vertebrate and ecdysozoan lineages[36–38] and cnidarian promoters as well (S10 Fig). By contrast, binding motifs for other factors involved in MCC differentiation, including Forkhead, E2f, and Myb, were also enriched at MCC gene promoters but this enrichment was less impressive, especially at MCC genes in other species (Fig 3A, S11 Fig).[20,22,39–41] These data support deep conservation of direct Rfx factor binding at the promoters of MCC genes but also raise the possibility that other factors are recruited to these promoters indirectly.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3.

Transcription factor motifs and binding in MCCs (A) Top de novo motifs identified in all MCC promoters in X. laevis along with the transcription factor family that best matches the motif, a p-value determined by the cumulative hypergeometric distribution, the frequency of the motif in the promoters analyzed, and the background frequency of the motif in all promoters. (B) The top de novo motif as in (A) that were found in promoters of all MCC paralogs in the indicated species. Hsa, Homo sapiens; Mmu, Mus musculus; Xla, Xenopus laevis; Dre, Danio rerio; Dme, Drosophila melanogaster; Nve, Nematostella vectensis. (C) Example screenshots around the promoters of three genes that were upregulated during MCC differentiation and one that was not (krt19) indicating H3K4me3 or transcription factor binding by ChIPseq. These promoters were bound by various combinations of Foxj1, Rfx2, Myb, and E2f4 as shown. (D) Shown are the top de novo motifs that are associated with all called Foxj1 ChIPseq peaks. (E) Shown are all core 950 MCC promoters and the combinations of E2f4, Foxj1, Rfx2, and Myb bound to each. Heatmap to the right indicates normalized expression counts in manipulations producing many MCCs (epithelial progenitors injected with Notch-icd compared to those injected with Notch-icd and Multicilin). (F) Shown is the change in expression between epithelial progenitors with few MCCs (injected with Notch-icd) versus progenitors with many MCCs (injected with Notch-icd and Multicilin) driven by all promoters bound by the indicated factors.

https://doi.org/10.1371/journal.pgen.1006538.g003

Transcription factor binding to MCC genes

To gain further insight into the TF binding required for MCC gene expression, we examined the binding targets of TFs recognizing the motifs enriched in MCC promoters: Rfx, Fox, E2f, and Myb. We interrogated recently published X. laevis ChiPseq data for Rfx2[22] and E2f4[40] and generated ChIPseq data for two additional candidate factors, Myb and Foxj1, using X. laevis epithelial progenitors manipulated to increase the number of motile cilia (see Methods, Fig 3C). We mapped reads from all datasets to the X. laevis genome (v9.1),[42] called peaks against input background, and subjected all peaks to de novo motif discovery. One of the top motifs identified in each set corresponded to the motif recognized by the factor immunoprecipitated, suggesting that these family members often bound the appropriate motif directly (Fig 3D, S12 Fig). However, this analysis also suggests possible co-binding; Foxj1 peaks at MCC genes were highly enriched for Rfx motifs, for example, whereas we saw no such enrichment across all promoters or enhancers (Fig 3D, S6 and S7 Figs).

We exploited the extensive binding data at promoters described above to assess which combination of TFs might account for the upregulation of MCC gene expression (Fig 3E). We found striking heterogeneity in the number of promoters bound by any one factor; likewise, we found heterogeneity in how often different combinations of factors bound to core MCC promoters. For example, Rfx2 bound a majority of these promoters (579 out of 693 MCC promoters bound by any of these factors), and other factors in this study rarely bound MCC promoters in its absence (114 out of 693, note extensive Rfx2 binding in Fig 3E). The largest overlap between two factors occurred between Foxj1 and Rfx2 (400), the next between E2f4 and Rfx2 (271), and finally the smallest overlap occurred between Myb and Rfx2 (230). A large fraction of core MCC promoters were not bound by any of these factors (257, or ~27%); however, we note that ~17% of MCC promoters were H3K4me3- (159/950), suggesting that their annotated TSS’s were incorrect.[43] Manual inspection of a subset of these revealed unannotated 5’ exons, and thus TSS’s, outside of our 2kb TSS window based on Mayball annotations (S13 Fig). These misannotations likely led to an underestimation of TF binding at the promoters of the core MCC gene list.

We asked whether binding of MCC transcription factors at core MCC promoters was associated with increased transcription, and examined transcriptional fold-changes of target genes between conditions repressing MCCs (Notch-icd) and conditions promoting MCCs (Notch-icd and Multicilin)(Fig 3F) (We did not examine enhancers with this approach: enhancers are more likely to influence transcription of nearby rather than distal genes[3] but often skip over the nearest TSS,[44] making target promoter prediction difficult). While Rfx2 alone bound many promoters across the genome, few of these genes exhibited increased expression in MCCs; neither did genes whose promoters were bound by E2f4 alone, Myb alone, or either co-bound with Rfx2 (Fig 3F, S14B and S15B Figs). By contrast, Foxj1 modestly increased expression of genes when bound to promoters alone and strongly increased expression when co-bound to promoters with Rfx2 (Fig 3F). While known to be involved in MCC differentiation,[41] Myb did not bind to many MCC promoters (Fig 3E), did not have a motif in MCC promoters nearly as often as Rfx or Forkhead proteins (Fig 3A), and was not nearly as associated with MCC gene transcription in promoters it did bind to (S14 Fig), suggesting Myb plays a more minor role in the differentiation of these cells. Other factors may also be involved, such as TP73; however, TP73 is not reported to strongly prefer MCC promoters (~3% of putative MCC genes in mouse),[45] nor was the TP73 consensus motif strongly enriched in X. laevis MCC promoters (Fig 3A), suggesting it may regulate more potent drivers of MCC genes such as Foxj1 rather than large numbers of MCC genes directly. Taken together, these genome-wide analyses largely support the current model that binding of both Foxj1 and a Rfx factor at a given promoter leads to potent activation of gene expression in MCCs while other combinations or individual factors were much less likely to do so.

Rfx2 influences Foxj1 binding

The coincidence of Foxj1 and Rfx2 binding at MCC genes was evaluated further by examining this binding in detail across the genome. We first estimated whether co-binding might occur simply by chance by evaluating the level of random overlap within areas of open chromatin (n = 1000 iterations), as defined by H3K4me3 and H3K27ac-positive regions. While random overlap of open chromatin predicts co-binding in 2208 peaks (SD 37), actual overlap between Rfx2 and Foxj1 binding occurred significantly more often (3505, p < 0.001, see Methods). We next evaluated where Foxj1 binding occurs in the genome in relation to Rfx2 binding. Foxj1 was rarely bound at promoters where Rfx2 binding was absent (~10%, 1203/11800 peaks, Fig 4A), but Foxj1 was often at promoters where Rfx2 binding also occurred (~42%, or 1487/3505 peaks). To determine if Rfx2/Foxj1 co-binding at promoters happens more often than by chance, we found that a random collection of Foxj1 and Rfx2 peaks (equal to the number of co-binding peaks, half Foxj1, half Rfx2) bound to promoters less often (~36%, or 1274 peaks, SD 27), suggesting that when cobound with Rfx2, Foxj1 was more likely to prefer promoters (p < 0.001, see Methods). Taken together, these data suggest that Foxj1 binding is strongly influenced by the present of Rfx2 binding, particularly at promoters.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4.

Interactions between Foxj1 and Rfx2 (A) Venn diagram at the top represents all peaks (fixed width) bound by Foxj1 and Rfx2 identified with ChIPseq, in terms of their overlap. The pie charts shown below illustrate where Foxj1 and Rfx2 binding occurs in relation to annotated genomic features depending on whether they bind together or alone. (B) Foxj1 ChIPseq peaks were subdivided into those located near promoters (< 1 kb to the TSS) or at more distal sites (> 1 kb to the TSS) and then analyzed for the frequency of Rfx or Fox binding motifs. (C) Shown is the distribution of Rfx2 and Foxj1 peaks relative to TAD boundaries (left panel), and the distribution of Rfx2 and Foxj1 sequencing tags relative to TAD boundaries (right panel).

https://doi.org/10.1371/journal.pgen.1006538.g004

We next compared the peaks of Foxj1 and Rfx2 binding to the location of TADs within the Xenopus genome. As a factor more likely to be found at distal sites, Foxj1 peaks were enriched inside domains (much like H3K27ac, Fig 2D), as compared to Rfx2 peaks (Fig 4C), which were enriched at their boundaries (similar to promoters, Fig 2C and 2D). When we examined the numbers of sequencing tags immunoprecipitated from these experiments, however, we saw a striking enrichment for both Rfx2 and Foxj1 at domain boundaries (Fig 4C), suggesting that while there were fewer Foxj1 peaks at TAD boundaries, those present were exceptionally strong. These results further indicate that Foxj1 binding at MCC genes located at TAD boundaries is associated with the presence of Rfx2 binding.

Rfx2 stabilizes Foxj1 on DNA

The simplest model to explain co-binding of Foxj1 and Rfx2 at MCC promoters is based on the presence of neighboring direct binding sites with appropriate consensus motifs, as shown in a handful of Drosophila cilia genes.[15] However, when we examined the enrichment of Forkhead and Rfx motifs in the promoters and distal sites bound by Foxj1 (Fig 4B, also see Fig 3D), we found that Rfx motifs were strongly present to an equal degree in promoters and distal sites, but that Forkhead motifs were depleted in promoters relative to distal sites. Moreover, when we performed de novo motif analysis on Foxj1 peaks not co-bound with Rfx2, we again saw robust enrichment of Rfx motifs, suggesting other Rfx proteins may also co-bind with Foxj1 (S17 Fig). Finally, despite Foxj1’s known conserved role in motile cilia formation,[20,35,39,46,47] Forkhead motifs were not particularly conserved in the promoters of MCC genes across vertebrate lineages (S11 Fig). Thus these findings suggest that the while Foxj1 is required to drive this process, consensus Forkhead motifs within promoters may be dispensable.

As there was a strong overlap between Foxj1 and Rfx2 peaks, and few Fox motifs at Foxj1-bound promoters (Fig 4B), we asked if Rfx2 might stabilize Foxj1 binding at these positions. Moreover, as many Foxj1 peaks were within TADs (Fig 4C), we speculated that Foxj1 at distal sites might be recruited to promoters, possibly at the TAD boundaries, in chromatin loops and that Rfx2 might facilitate this recruitment. To directly test the possibility that Rfx2 stabilized Foxj1 binding, we knocked down Rfx2 with a well-characterized morpholino[22,48] and performed ChIPseq on Foxj1 (Fig 5). Out of 15,305 Foxj1 peaks identified in control tissue, some 6,360 were reduced at least 3-fold (~42%), and 2,084 were reduced 10-fold (~14%) in an Rfx2 knockdown. We saw strong reductions at key promoters, such as the rfx2 promoter itself (Fig 5A), while other promoters maintained Foxj1 occupancy (Fig 5B). We also saw strong reductions in regions with many peaks, such as the superenhancer surrounding tubb2b (Fig 5C); in cases where reduced Foxj1 peaks did not overlap with robust Rfx2 peaks we often saw Rfx motifs (Fig 5C) and we also found Rfx motifs enriched in Foxj1 peaks not cobound by Rfx2 (S17 Fig), further hinting that other Rfx proteins might be involved. When we looked across all regions bound by Foxj1, we saw a reduction of Foxj1 tags at those positions in the absence of Rfx2 and striking reduction of Foxj1 tags at the promoters of MCC genes (Fig 5D). By contrast, we saw a slight increase at non-MCC promoters, which may be the result of excess Foxj1 protein binding nonspecifically to regions of open chromatin.[49,50] While there is a large fraction of Foxj1/Rfx2 cobound sites in promoters that are not upstream of genes in our core MCC list (3025 out of 3505), we found the greatest Foxj1 reduction at core MCC promoters (S18 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5.

Binding of Foxj1 at promoters is dependent on Rfx2 (A-B) Screenshot around the promoter of X. laevis rfx2 (A) or hdx (B) along with the sequence tags obtained in ChIPseq of H3K4me3, H3K27ac, Foxj1, or Rfx2 in wildtype and Foxj1 in Rfx2 morphants (MO). (C) Shown is a screenshot of the X. laevis genomic region containing the tubb2b gene with ChIPseq tracks as in (A). The position of all Rfx motifs is denoted in the bottom track. Foxj1 peaks that are reduced >3-fold in the Rfx2 morphants compared to control are shaded. (D) Shown are sequencing tag histograms in peaks as labeled from ChIPseq of Foxj1 in wild-type progenitors or progenitors from Rfx2 morphants.

https://doi.org/10.1371/journal.pgen.1006538.g005

Rfx2 might stabilize Foxj1 at sites where Foxj1 binds directly to its consensus sequence in DNA, or alternatively, it might stabilize Foxj1 at positions where Foxj1 binds indirectly. To distinguish between these possibilities, we examined changes in Foxj1 binding in peaks occupying positions containing either a strong Forkhead motif but not a Rfx motif (2,646 peaks) or a strong Rfx motif but not a Forkhead motif (2,083 peaks). We found that while tag counts in positions containing a Forkhead motif alone were modestly reduced in the absence of Rfx2, Foxj1 tag counts at positions with an Rfx motif alone were drastically reduced (S18 Fig), suggesting that Rfx2 might have a minor role in stabilizing Foxj1 directly but a much larger role in stabilizing it indirectly.

Rfx2 stabilizes Foxj1 inside chromatin loops

The increased likelihood of transcription of promoters bound by multiple factors is consistent with general findings by others[3,11] but does not explain the distribution of Foxj1, Rfx2, and their respective motifs relative to transcriptional start sites. Rather, the results above suggest a model where, in some cases, Rfx proteins bind at both promoters and distal enhancers, and through dimerization,[51,52] maneuver the two together to effect transcription. We further speculated that this looping enables Foxj1, bound to enhancers, to localize to promoters, where it was then crosslinked and immunoprecipitated in our assays. This proposed looping would create a protein-DNA complex wherein Foxj1 bound enhancer DNA through a motif but was held nearby its target promoter DNA via Rfx2 proteins: ChIP of a single Foxj1 protein would then pull down both its directly-bound enhancer and its indirectly-bound promoter. To evaluate this model, we first determined all significant interactions in our 3D chromosomal conformation data from wild-type epithelial progenitors using an approach published previously (see Methods, “TCC significant interactions and TADs”).[25] We then interrogated the anchor points of these interactions for co-binding of Rad21, surmising that the anchor points of significant chromosomal interactions would be enriched for known looping machinery.[28,53] To test the probability of overlap between positions bound by Rad21 and the anchor endpoints of significant interactions, we performed a hypergeometric distribution test, calculating the amount of overlap by chance alone, the amount of enrichment over this amount measured in our data (if any), and the probability of encountering this enrichment.[3,54] Pairs of Rad21 peaks were found strongly enriched at the endpoints of loop anchors (Rad21 self-interacting loops in Fig 6A), in agreement with [53]. We then applied this approach to other peaksets, finding strong enrichment for pairs of Rfx2 peaks at the endpoints of looping interactions, consistent with our model that Rfx2 could mediate such interactions. We also saw pairs of Foxj1 enriched at the ends of loops, as well as Rfx2-dependent Foxj1 peak pairs (labeled as “F3”, for peaks reduced by 3-fold or more in the absence of Rfx2). There was also a strong enrichment for pairs of MCC promoters at the endpoints of loops. Finally, we asked if the peaksets enriched at anchor endpoints changed during MCC differentiation, by comparing wild-type interaction data with data obtained with progenitors expressing Multicilin (Fig 6B). We saw an increase in enrichment of both Foxj1 peaks and Rfx2-dependent Foxj1 peaks at anchor endpoints, and an even stronger enrichment of endpoints with Rfx2-dependent Foxj1 peaks at one end and MCC promoters at the other end (connection between “F3” and “MCC”). Taken together, these data suggest a model wherein Foxj1 peaks bound to distal enhancers are recruited to MCC promoters via Rfx2 dimerization, a process that coordinates or stabilizes chromatin loops (Fig 6C and 6D).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6.

Chromatin loops connect MCC regulatory elements (A) The relative enrichment over expected of histone modifications or transcription factor binding sites at loop anchor points was calculated and visualized in Cytoscape. “Wild-type” tissue is unmanipulated progenitors containing a mixture of outer cells, ionocytes, and multiciliated cells. Expected overlap was determined by hypergeometric distribution; 3D interactions were obtained from wild-type progenitors, and line thickness is inversely proportional to p value (range: 1e-37 to 1e-611, thicker line is lower p value). Nodes are as labeled; “F3” represents the subset of Foxj1 ChIPseq peaks that are reduced 3-fold or greater in Rfx2 knockdowns and “MCC” represents MCC TSS’s. (B) 3D interactions were obtained for wild-type progenitors using progenitors injected with Multicilin to increase numbers of MCCs as background (to determine interactions stronger in wild-type tissue) and 3D interactions were also obtained using the reverse (to determine interactions stronger in multiciliated cells). Relative enrichments of histone modifications or transcription factor binding sites were determined for each as in (A) and then compared to one another. Thus, values here depicted by color represent changes in enrichment between the two conditions. (C,D) Model of recruitment of Foxj1-bound enhancers to MCC promoters via Rfx2 dimerization. (E) Model of how Rfx2-mediated enhancer recruitment operates in the context of TAD boundaries.

https://doi.org/10.1371/journal.pgen.1006538.g006

Ethics statement

All animal experiments reported here were carried out at the Salk Institute, supervised by a licensed veterinarian in the Animal Resource Department, under guidelines that are AAALAC certified, and using procedures reviewed and approved by the Salk IACUC committee under protocol #15–00034.

Embryos, injection, and isolation of epithelial progenitors

X. laevis embryos were prepared by in vitro fertilization using standard protocols.[66] Synthetic, capped mRNA was generated in vitro using DNA templates described below and injected into embryos at the two-cell or four-cell stage (typically 0.1–5.0 ng of RNA per embryo), targeting the four animal quadrants. Pluripotent epidermal progenitors (animal caps) were isolated at stage 9–10 and cultured in 0.5X MMR and harvested at 3, 6, and 9 hours at 22 deg after mid-stage 11. Progenitors injected with RNA encoding HGR-inducible constructs (Multicilin-HGR and Foxi1-HGR) were induced with 1μM dexamethasone at mid-stage 11.

DNA constructs and knockdowns

DNA templates for synthesis of mRNAs have been described previously for Notch-icd, dbm, Foxi1-HGR, Multicilin-HGR, and dominant negative Multicilin.[18,19,67] The DNA template for generating an RNA encoding a tagged form of Foxj1 was obtained by subcloning a cDNA encoding Foxj1[39] into a CS2 vector with a FLAG tag, while that for a tagged form of Myb was generated by PCR amplification of a myb cDNA from a stage 17 cDNA library and subcloning into a CS2 vector with a GFP tag. The Rfx2 morpholino[22,48] was a kind gift of Meii Chung and John Wallingford.

RNAseq libraries

RNAs were isolated by the proteinase K method followed by phenol-chloroform extractions, lithium precipitation, and treatment with RNase-free DNase and a second series of phenol-chloroform extractions and ethanol precipitation. RNAseq libraries were then constructed (Illumina TruSeq v2) and sequenced on an Illumina platform. Details on specific experiments are in S17 Table, and RNAseq reads are deposited at NCBI (GSE76342).

RNAseq informatics

Sequenced reads from this study or obtained previously[22,40] were aligned to the X. laevis transcriptome, MayBall version[22] with RNA-STAR[68] and then counted with eXpress.[69] Effective counts from eXpress were then clustered (Cluster 3.0[70], log transformed, maxval—minval = 6) and visualized with Java Treeview (v1.1.6r, [71]) to produce heatmaps. DESeq was used to estimate dispersion and test differential expression using rounded effective counts from eXpress.[72] Changes in expression were visualized in R with beanplot (https://cran.r-project.org/web/packages/beanplot/beanplot.pdf), and to visualize RNAseq reads in a genomic context they were mapped to genome version 9.1 with bwa mem[73] and loaded as bigWig tracks into the Integrative Genomics Viewer browser.[74]

Multispecies comparisons

Orthologs/paralogs of MCC genes were obtained from gene assignments in Homologene or by using official gene symbols in N. vectensis gene models.[75] Promoter collections for H. sapiens, M. musculus, D. rerio, and D. melanogaster were obtained from UCSC. To determine N. vectensis promoters, 5’ ESTs were mapped to JGI genome build v1.0 with BLAT.[76] H3K4me3 ChIPseq reads from N. vectensis[77] were then mapped to v1.0, peaks called, and then overlapped with TSS’s as determined by 5’ ESTs to verify promoters, similar to our approach in confirming X. laevis promoters. 5’ ESTs were then aligned to N. vectensis gene models (v1.0) to connect transcriptional start sites to gene identity. Motif enrichment on all promoters and MCC ortholog promoters was done with HOMER.[3] Species divergence times are from Timetree.[78]

ChIPseq libraries

Because ChIP-grade antibodies are generally not available that recognize Xenopus proteins, we tagged Myb and Foxj1 with GFP and FLAG, respectively, and injected mRNAs encoding these constructs into embryos. Myb has functions in stem cell maintenance independent of a role in MCCs,[79] so to enrich for motile cilia targets, we performed ChIPseq on Myb in progenitors injected with Multicilin-HGR induced at mid-stage 11 (our previous work with E2f4 ChIPseq[40] also included Multicilin injections, allowing for direct comparison of MCC-enriched TF targets). Moreover, overexpression of Foxj1 promotes ectopic cilia[39], so samples injected with that construct were also enriched for motile cilia targets by virtue of expressing a tagged Foxj1 construct.

Samples were prepared for ChIP using described methods[40] with the following modifications: About 250 animal caps for TFs or 100 caps for histone modifications were fixed for 30 min in 1% formaldehyde, and chromatin was sheared on a BioRuptor (30 min; 30 sec on and 2 min off at highest power setting). Tagged proteins with associated chromatin were immunoprecipitated with antibodies directed against GFP (Invitrogen catalog no. A11122, lot no. 1296649) or FLAG (Sigma, cat #F1804). Native proteins were immunoprecipitated with antibodies directed against H3K4me3 (Active Motif, cat #39159; lot #01609004), H3K27ac (Abcam, cat #ab4729, lot #GR71158-2), or rad21 (Abcam, cat #ab992; commercially-available CTCF antibodies target regions not conserved in the X. laevis protein). DNA fragments were then polished (New England Biolabs, end repair module), adenylated (New England Biolabs, Klenow fragment 3′–5′ exo- and da-tailing buffer), ligated to standard Illumina indexed adapters (TruSeq version 2), PCR-amplified (New England Biolabs, Phusion or Q5, 16 cycles), and sequenced on an Illumina platform. Details on specific experiments are in S17 Table, and ChIPseq reads are deposited at NCBI (GSE78176).

ChIPseq informatics

ChIP-seq reads from this study or published previously[22,40] were mapped to X. laevis genome v9.1 with bwa mem,[73] peaks called with HOMER[3] using input as background. Peak positions were annotated relative to known exons (Mayball gene models[22]), with promoters defined as being +/- 1Kb around the TSS. Peak sequences were interrogated for de novo motif enrichment with HOMER and MCC promoters were clustered (based on if they were bound/not bound) with Cluster 3.0 and visualized with Java Treeview (v1.1.6r). Tags or motifs were counted at peak positions with HOMER and plotted with Excel or R. To determine enrichment statistics, we used the hypergeometric distribution to evaluate the likelihood of peak overlap.[3] Additionally, as regions of open chromatin occupy a fraction of the entire genome, we also assessed overlap between pairs of factors by determining the distribution of randomly picked open chromatin sites (equal to the number of sites in the first and second TFs in the comparison; open chromatin was estimated by using all H3K4me3 and H3K27ac sites[80]) and repeated this approach 1000 times to get a distribution of overlap between two sets of random regions of open chromatin. We then compared this distribution to the overlap observed in pairs of measured TFs, again restricted to open chromatin (as in Fig 4 and S14 and S15 Figs). To determine if subsets of overlapped peaks preferred promoters or other features more than by chance, we employed a similar approach: instead of selecting overlapped peaks, we sampled from all peaks, half from each set, and determined how many peaks form this null set bound to promoters. We repeated this operation 1000 times to obtain the distribution of promoter preference from these random combinations. Finally, when determining differential binding and generating bigwig files for visualization, each experiment was normalized to 10M reads in HOMER-style tag directories to account for differences in sequencing depth.

Tethered conformation capture

TCC was performed as described[81] with the following modifications: 100 animal caps were harvested at the 9 hour timepoint and fixed for 30 minutes in 1% formaldehyde. Tissue was thawed, rinsed in PBS, and lysed directly in 250 μl ice-cold wash buffer with SDS added (50 mM Tris.HCl pH = 8.0, 50 mM NaCl, 1 mM EDTA, 0.56% SDS) rather than using lysis buffer. A complete protocol is available on request. Details on specific experiments are in S17 Table, and TCC reads are deposited at NCBI (GSE76363).

TCC informatics

TCC reads were cleared of PCR duplicates, trimmed down to exclude MboI restriction sites (GATC) and mapped to the X. laevis genome v9.1 with bwa mem. Topological domain peakfiles, interaction matrices, and metagene plots were generated with HOMER (additional details below). Matrices were visualized with Java TreeView.

TCC background model

All TADs and significant interactions were called against a background of expected interactions. This background model was constructed using the approach described in [25]. Briefly, the frequency of proximity-ligated fragments that are connected across a certain distance on a linear chromosome is expected to increase as that distance shrinks. To generate the model, the genome is broken into bins with a sliding-window approach to boost signal, and sequence depth in those regions is normalized to account for mapping errors and biases in restriction enzyme accessibility. Once depth is normalized in all regions, the expected variation in interaction between regions is calculated as a function of distance.

TCC significant interactions and TADs

As the background model estimates how many interactions between regions are expected given a specific linear distance and sequencing depth, we looked at the number of interactions we actually observed between regions in the TCC data. To determine interactions with some measure of statistical confidence, we used the cumulative binomial distribution, where the number of trials was the total number of reads mapping within regions, success was the expected interaction frequency, and observed successes came directly from read pairs connecting regions from the TCC data itself. We then used the hypergeometric distribution to ask if the anchor points of these interactions overlapped with ChIPseq peaks or other genomic features more than would be expected by chance (Fig 6A) and visualized these values with Cytoscape.[82] We also determined interactions that were stronger in wild-type or Multicilin-injected epithelial progenitors: after determining significant interactions for one of these conditions, we used TCC reads from the other to score their strength (Fig 6B). TADs were determined by calculating a directionality index with HOMER; we observed similar numbers of TADs with varying resolutions (10–25 kb). All methods in this section were adopted from [25].