Gene silencing dynamics are modulated by transiently active regulatory elements

SUMMARY

Transcriptional enhancers have been extensively characterized, but cis-regulatory elements involved in acute gene repression have received less attention. Transcription factor GATA1 promotes erythroid differentiation by activating and repressing distinct gene sets. Here, we study the mechanism by which GATA1 silences the proliferative gene Kit during murine erythroid maturation, and define stages from initial loss of activation to heterochromatinization. We find that GATA1 inactivates a potent upstream enhancer but concomitantly creates a discrete intronic regulatory region marked by H3K27ac, short noncoding RNAs, and de-novo chromatin looping. This enhancer-like element forms transiently and serves to delay Kit silencing. The element is ultimately erased via the FOG1/NuRD deacetylase complex, as revealed by study of a disease-associated GATA1 variant. Hence, regulatory sites can be self-limiting by dynamic co-factor usage. Genome-wide analyses across cell types and species uncover transiently active elements at numerous genes during repression, suggesting that modulation of silencing kinetics is widespread.

E-toc

• Time-resolved analyses reveal transient enhancer elements during gene silencing

• A GATA1-driven transient enhancer delays GATA1-mediated silencing of the Kit gene

• FOG1/NuRD is required for the self-limitation of GATA1-occupied transient elements

• Transient, potential silencing-delay elements seem to be widespread

E-blurb

Transcription factor GATA1 can activate and repress transcription. Vermunt et al. show that GATA1-induced silening of the Kit gene is associated with GATA1-induced formation of a transient enhancer that delays Kit silencing. Co-factor usage determines the self-limiting nature of GATA1-occupied transient regulatory elements.

Graphical Abstract

INTRODUCTION

Terminal cellular differentiation requires coordination of gene activation and repression. Cell type-characteristic genes are upregulated while those that control proliferation are often downregulated. Hallmarks of transcriptional activation are changes in histone modifications as well as gained enhancer-promoter loops^1,2. Even though chromatin looping has been associated with gene repression^3–8, its role during this process is much less clear. Gene silencing can be achieved by various means. First, it could occur through downregulation of activators as might happen during developmental state transitions (operationally defined here as passive silencing). Second, it could be an active process such as the direct displacement of transcriptional activators by repressors (operationally defined here as active silencing), and/or the exchange of active for repressive histone modifications. An example for active repression is provided by the substitution of transcription factor GATA1 for GATA2 at cis-regulatory elements during erythroid differentiation which leads to downregulation of numerous genes⁹. After initial loss of activation, a repressed state may be maintained via heterochromatinization in the absence of the silencing initiating factors¹⁰.

Gene regulation by the erythroid transcription factor GATA1 offers an ideal model to study these processes because GATA1 activates lineage-specific genes while at the same time directly repressing genes associated with the progenitor state, such as the gene encoding the cytokine receptor Kit. Mutations in GATA1 can variably affect these functions, causing anemia or malignancies¹¹. The silencing of Kit transcription by GATA1 is accompanied by architectural re-organization of the Kit locus in murine and human cells^4,12,13, but the dynamics of histone modifications and fine-scale architectural features have not been characterized. Here, we exploit a system that relies on conditional activation of GATA1 to explore stages of Kit gene repression, and investigate the contribution and dynamics of the different regulatory elements involved at high temporal resolution. We discover that transcriptional silencing, which is associated with GATA1-mediated loss of a strong upstream enhancer, does not follow a simple linear path, but is delayed by GATA1-driven activating elements that appear transiently during the silencing process. Hence, GATA1 has a dual role within the Kit locus thereby fine-tuning the kinetics of gene downregulation. Genome-wide analyses of enhancer signatures during erythroid differentiation as well as human adipocyte differentiation confirm the existence of transient enhancer elements near numerous downregulated genes. Therefore, transient elements may buffer against the rapid loss of enhancer activity to fine-tune transcriptional silencing during cellular differentiation in a variety of tissues.

RESULTS

Regulatory elements with enhancer-like features transiently arise during Kit gene silencing

To dissect the dynamics of Kit transcriptional silencing by GATA1, we employed G1E and G1E-ER4 erythroblast cell lines¹⁴. G1E cells lack GATA1 and proliferate in an immature state. G1E-ER4 cells, derived from G1E, stably express GATA1 fused to the ligand binding domain of the estrogen receptor (ER). Exposure of G1E-ER4 cells to estradiol (E2) activates GATA1, which in turn triggers global changes in gene expression, such as the upregulation of erythroid genes, including the globin genes, and repression of genes associated with the proliferative state, such as Kit. This well-characterized system faithfully recapitulates normal erythroid maturation, including transcriptional changes, histone modifications, and chromatin architecture dynamics^12,15–18. The added advantage of this system is that cells behave in a highly synchronous manner, facilitating the measurements of transient features that might escape detection in less homogenous populations. G1E-ER4 cells represent a slightly more differentiated state compared to parental G1E cells because even in the absence of E2, basal level (“leaky”) GATA1 chromatin occupancy is observed⁴. Kit is highly transcribed in erythroid precursors, including G1E cells, driven by an enhancer located 114 kb upstream (−114 kb) of the transcription start site (TSS) that is in physical proximity with the promoter^4,19. Kit is the only gene within its topological associating domain²⁰ (Figure S1A), which facilitates studies of chromatin and architectural features due to the absence of potentially confounding features from other genes.

We first assessed Kit repression dynamics by measuring by RT-qPCR Kit primary transcripts, which unlike mature mRNAs are not influenced by RNA stability. Untreated G1E-ER4 cells expressed lower levels of Kit compared to G1E cells, a likely reflection of the described basal GATA1-ER chromatin occupancy and activity in the absence of E2. At 13 hours (hr) of E2 exposure, Kit expression was significantly reduced, and at 24 hr became almost undetectable (Figure 1A). To determine whether GATA1-ER decommissioned positive Kit regulatory elements, we profiled via ChIP-seq the activation-associated mark H3K27ac (acetylated histone 3 lysine 27) under the same conditions. Commensurate with the level of Kit transcription, H3K27ac was diminished across the locus in untreated G1E-ER4 cells compared to G1E (Figure 1B). GATA1-ER activation triggered a rapid and dramatic loss of H3K27ac at the −114 kb enhancer and promoter, suggesting that loss of enhancer activity accounts to a large extent for Kit silencing by GATA1. However, surprisingly, H3K27ac was gained at intronic GATA1-occupied regulatory elements at +5 kb, +33 kb and +49 kb with respect to the TSS, with the highest levels occurring at 13 hr, a time at which Kit repression had already markedly progressed (Figure 1B, Figure S1B). H3K27ac levels were much less pronounced and not dynamic at the +58 kb GATA1 occupied site, the reason for which is unknown. At 24 hr, all GATA1-occupied intronic elements were less enriched for H3K27ac than at 13 hr but remained detectable.

Figure 1. Intronic GATA1 binding sites transiently gain H3K27ac and eRNAs during Kit gene silencing.

A) Upper panel: cell lines and time points used in this study. Lower panel: Kit primary transcript (pT) levels normalized to Gapdh and relative to the mean of G1E-ER4 -E2 replicates. n=3 for G1E, n=4 for all other time points. Error bars represent standard error of the mean (SEM). One-way ANOVA: ****P < 0.0001. B) Anti-H3K27ac ChIP-seq tracks for G1E and G1E-ER4 -E2, +E2 13 hr and +E2 24 hr in green. n=2 for each condition, average of replicates depicted. Anti-GATA1 ChIP-seq track for G1E-ER4 +E2 13hr from Jain et al.¹⁷ in red. GATA1-ER occupied sites of interest are highlighted by boxes. C) PRO-seq tracks for the Kit locus. n=3 for G1E-ER4 -E2 and n=2 for G1E, G1E-ER4 +E2 13 and 24 hr, average of replicates depicted. The insets on the right show enlarged images of boxed GATA1 sites of interest. Red: plus strand; blue: minus strand. Related to Figure S1.

The unexpected gain in H3K27ac at intronic GATA1 elements during repression prompted us to consider whether they might represent transient enhancer-like elements. Since active enhancers are marked by eRNAs^21,22, we addressed this question using PRO-seq (precision nuclear run-on sequencing) which measures RNA Polymerase II-mediated nascent transcription in a stranded manner at single base resolution²³. We compared the PRO-seq signal dynamics to publicly available datasets of mature RNA in G1E-ER4 cells at 0 (-E2) , 13 and 24 hr of differentiation (+E2)¹⁷. Relative changes between time points were globally correlated between the two data types (-E2 versus +E2 13 hr, r=0.754 and +E2 13 versus 24 hr, r= 0.591; Figure S1C). Remarkably, even though sense Kit transcription was diminished upon 13 hr of E2 treatment, short antisense transcripts were gained transiently at the GATA1-occupied, H3K27ac positive regions +5 kb, +33 kb, and +49 kb (Figure 1C). Sense oriented eRNAs would be obscured by Kit pre-mRNA. The antisense transcripts disappeared again towards 24 hr concomitantly with reduction of H3K27ac enrichment. This suggests that loss of activation by the −114 kb enhancer is accompanied by GATA1-mediated formation of transient enhancer-like elements during the process of transcriptional repression.

Transient regulatory elements form de novo chromatin contacts during Kit silencing

Since regulatory elements are frequently in physical proximity with their target genes, we performed 4C-seq to analyze the 3D configuration of the Kit locus at various stages of repression. PeakC was used to identify significant interactions²⁴. In line with previous observations⁴, the activation-associated −114 kb enhancer-promoter loop was diminished at 13 hr and essentially absent at 24 hr of E2 treatment (Figure 2A and Figure S1D). In contrast, 4C-seq contacts between the promoter-proximal region and the +49 kb element increased during repression (Figure 2A and Figure S1D). Reciprocal 4C-seq using +49 kb as a viewpoint revealed the +5 kb element as its most prominent interaction site (Figure S1E). Indeed, 4C data for the +5 kb viewpoint showed a strong loop with +49 kb at 13 hr and 24 hr (Figure 2B,C). Since contacts between the +49 kb element and the Kit promoter were also gained at 13 hr and 24 hr (Figure 2A and Figure S1D), it is unclear whether the gained contacts of the +49 kb element involve the promoter and the +5 kb region separately or in combination. Either scenario is possible since enhancer elements can not only contact promoters but also other regulatory elements. In concert, these results show that during the process of transcriptional repression, transient enhancer-like elements emerge that are capable of forming dynamic long-range chromatin contacts.

Figure 2. Transient +49 kb element can delay Kit gene silencing and prevent heterochromatinization.

A) 4C-seq in G1E (purple) and G1E-ER4 +E2 13 hr (orange) using the Kit promoter as a viewpoint. Tracks are plotted on top of each other with the lowest one for each DpnII fragment depicted in grey. Significance of chromatin interactions was calculated using peakC (settings qWd:1.5, qWr:1.25, asterisk = significant interaction). B) Same as in (A) for G1E (purple) and G1E-ER4 +E2 13 hr (orange) using Kit +5 kb as a viewpoint. (C) Same as in (B) for G1E (purple) and G1E-ER4 +E2 24 hr (red). D) Kit primary transcript (pT) levels up to 24 hr in unedited control (n=3) and Δ+49 kb (n=3) subclones of clone A, normalized to Gapdh and relative to −E2 for each subclone. Error bars represent SEM. An unpaired, two-tailed t-test was done for each time point: P values indicated, * P < 0.05, ** P < 0.01. E) Same as in (D) for clone B. F) Same as in (D) for clone C. G) 4C-seq in clone A control subclone 1 at time points -E2 (yellow) and +E2 13 hr (orange) using the Kit +5 kb as a viewpoint. H) Same as in (G) for clone B control subclone 1. I) Same as in (G) for clone C control subclone 1. J) Left, schematic of the experimental set up. Right, Kit pT levels normalized to Gapdh and relative to the mean of G1E-ER4k control clones for G1E-ER4k clones (n=9), −114 kb enhancer deletion clones (n=11) and +49 kb with subsequent −114 kb enhancer deletion clones (n=6). Error bars represent SEM. K) Anti-H3K27ac (green) and anti-H3K27me3 (dark blue) ChIP-seq tracks. G1E-ER4k cells before and after 24 hr differentiation, two Δ−114 kb ~10% Kit clones and two Δ−114 kb 0% Kit clones are depicted. GATA1 sites of interest are highlighted by boxes. Exons are covered by a transparent white box to mask H3K27ac signal resulting from Kit cDNA overexpression. L) 4C-seq for G1E-ER4k (dark grey) and Δ−114 kb 10% clone 1 (yellow) using the Kit +5 kb site as a viewpoint. Related to Figure S2.

Transient enhancer-like element delays Kit transcriptional silencing

De novo chromatin contacts at the Kit locus during transcriptional silencing had suggested that they are repressive in nature and might function by physically sequestering the promoter away from the −114 kb enhancer⁴. However, the gains in H3K27ac and eRNAs suggest an additional, if not alternative, model in which the gained contacts enhance transcription transiently. To test this idea directly, we deleted via CRISPR/Cas9 a 190 base pair (bp) segment at the +49 kb element (Δ+49 kb) bearing three GATA1 consensus motifs (Figure S1F). The reason for focusing on the +49 kb element instead of other elements is that this element appears from a low ground state, exhibits the highest relative increase in H3K27ac between -E2 and 13 hr, as well as a significant reduction of H3K27ac upon completion of repression between 13 and 24 hr (Figure S1B), thus representing a truly transient element. Moreover, unlike another transient element at +33 kb, the +49 kb element engages in a long-range chromatin contact with the promoter-proximal region making it a strong candidate for Kit regulation.

We generated three unedited control subclones and three Δ+49 kb subclones from three G1E-ER4 clones A, B and C (Figure S1F). Δ+49 kb did not affect Kit expression levels in undifferentiated cells (Figure S1G) and H3K27ac enrichment was no longer gained at 13 hr, indicating that the deletion eliminates formation of the +49 kb element (Figure S1H). 4C-seq analysis revealed that loop formation at 13 hr was completely abolished in Δ+49 kb subclones (Figure S1I), further corroborating that the +49 kb element contains the intronic loop anchor. RT-qPCR analysis during 24 hr after GATA1-ER activation showed that all Δ+49 kb subclones were still capable of Kit repression (Figure 2D–F). Primers against eRNAs from the −114 kb enhancer confirmed that in all subclones (both control and Δ+49 kb), enhancer activity is rapidly lost after GATA1 activation, further suggesting that enhancer decommissioning drives loss of Kit gene expression (Figure S1J).

Importantly, in clone A- and C- derived subclones Kit gene silencing was significantly accelerated in Δ+49 kb subclones. In the clone B-derived lines, we observed a trend towards faster silencing but effect sizes were more modest (Figure 2D–F). Myc, another proliferative gene, silencing was not significantly altered in any of the clones, showing the effect of Δ+49 kb is Kit gene specific (Figure S1K). 4C-seq at the Kit locus revealed a clear +5 kb/+49 kb chromatin loop at 13 hr in clones A and C but much less so in clone B (Figure 2G–I). This suggests a correlation between loop formation and the silencing delay effect of the +49 kb element. In conclusion, rather than having a role in repression, GATA1-ER binding at intronic sites creates activating elements that can counteract gene downregulation.

Intronic cis-regulatory Kit element can prevent heterochromatinization

Completion of Kit repression in the absence of the +49 kb element suggests that GATA1-ER-mediated inactivation of the −114 kb enhancer is the major mechanism underlying Kit downregulation, with the +49 kb element playing a modulatory role. To test whether the +49 kb element indeed modulates repression in that it can transiently buffer against loss of −114 kb, we examined the effects of deleting the −114 kb enhancer alone or in combination with the +49 kb element using CRISPR/Cas9 in G1E-ER4 cells (Figure 2J). Because in their undifferentiated state these cells rely on Kit signaling for proliferation and viability, we first generated G1E-ER4 cells that constitutively express the Kit cDNA (G1E-ER4k cells). Forced Kit expression did not affect the levels of endogenous nascent Kit transcripts as assessed by RT-qPCR using an intron-specific primer pair (Figure S2A). A 200 bp deletion spanning the −114 kb enhancer markedly diminished Kit expression in three independent clones. We narrowed down the region by deleting only 100 bp, and ultimately perturbing just a single consensus GATA element within the enhancer and found the effects to be comparable (Figure S2B–D). The GATA element, which is bound by GATA2 in immature erythroblasts⁴ is thus essential for Kit enhancer activity. On average, primary Kit transcript levels decreased by approximately 85% after −114 kb enhancer disruption (Figure 2J and Figure S2B), but marked variation was observed between clones with some having no Kit expression at all, and others maintaining Kit levels between 10% and 20% (Figure 2J and Figure S2B).

The clonal variability in expression might be accounted for by residual enhancer activity at the locus because enhancers can stimulate transcription in a variegated manner that can be propagated over cell generations²⁵. We hypothesized that the residual Kit transcriptional activity could be driven by the +49 kb element. This element is inactive in parental GATA1 null G1E cells, but is marked by low levels of H3K27ac in untreated G1E-ER4 and G1E-ER4k cells (Figure 1B and 2K), presumably due to basal GATA1-ER activity. Indeed, compared to G1E cells, G1E-ER4 cells displayed measurable GATA1-ER binding at the +49 kb element (Figure S2E). When the +49 kb element (100 bp deletion) was deleted in addition to the −114 kb enhancer (100 bp deletion), six out of six clones lost Kit transcription entirely (Figure 2J). This is in agreement with the +49 kb element functioning as an enhancer that can variably keep the Kit locus transcriptionally active in the absence of the −114 kb enhancer.

To elucidate the difference between Δ−114 kb clones with and without residual transcription, we profiled H3K27ac in two clones that sustained ~10% Kit transcription and two clones in which the Kit gene was silenced following disruption of the −114 kb GATA motif (Figure S2C,D). The clones with ~10% Kit transcription maintained H3K27ac at the promoter proximal region and intronic GATA sites, including the +49 kb element. In contrast, the silent clones were devoid of H3K27ac at all elements (ChIP-seq tracks Figure 2K and RT-qPCR Figure S2F,G). To assess whether the +49 kb element functions as a weak enhancer by forming contacts with the promoter proximal region, we carried out 4C-seq. In the Kit expressing Δ−114 kb clones contacts between +49 kb and the promoter proximal +5 kb region were higher than in control G1E-ER4k cells (Figure 2L and Figure S2H).

The complete loss of Kit transcription in a subset of Δ−114 kb clones prompted us to examine whether this was due to the acquisition of H3K27me3, a mark indicative of facultative heterochromatin. We found that silent (0%) clones, but not the ~10% expressing clones, gained H3K27me3 at the promoter and surrounding region (Figure 2K and Figure S2G). Therefore, the +49 kb element can prevent heterochromatinization in a variegated manner. In line with general compaction and inactive intronic elements, no chromatin loop was observed between +5 kb and +49 kb in silent clones (Figure S2I). This result suggests a model in which acute GATA1-ER induced silencing leads to loss of H3K27ac and diminished enhancer-promoter contacts, while prolonged gene inactivation, such as that observed during clonal selection of Δ−114 kb clones, cements a permanent silent state via heterochromatin formation. Consistent with such a dynamic model, acute GATA1-ER-induced Kit repression in G1E-ER4 cells failed to achieve H3K27me3 enrichment (as low level H3K27ac remained at transient intronic regulatory sites), but H3K27me3 acquisition is the in vivo end point of Kit gene silencing in murine mature red blood cells (Figure S2J)^17,26.

Genome-wide analysis identifies hundreds of transiently active bidirectionally transcribed regions

The PRO-seq data enabled assessment of whether transient enhancer-like elements appear at other loci during their silencing. Our approach was to first classify all genes according to their dynamics of nascent transcription during GATA1-mediated differentiation. PCA analysis showed clustering of replicates at each time point, consistent with high replicate concordance, and clear separation of the time points (Figure S2K). Active genes with a significant change (ANOVA, P < 0.01) across the four timepoints were selected and categorized into 16 groups by unsupervised clustering (Figure 3A and Table S1). Transcriptional dynamics for each cluster were comparable to previous reports of mature RNA in this system (Figure S2L).

Figure 3. Genome-wide analysis reveals many more transient regulatory elements at downregulated genes.

A) Genes with significant ANOVA results across timepoints grouped using unsupervised hierarchical clustering. Colors correspond to the average normalized expression of replicates for each time point (number of standard deviations from G1E) with blue: lower and red: higher. B) Same as in (A) for bidirectionally transcribed dREG peaks (non-TSS sites only). C) PRO-seq tracks for the Myc and Mitoferrin enhancers as well as a newly identified transient intronic site for Prmt3. n=3 for G1E-ER4 -E2 and n=2 for G1E, G1E-ER4 +E2 13 and 24 hr, average of replicates depicted. D) Same as in (C) for the Atp1a1 locus. E) 4C-seq for G1E (purple) and G1E-ER4 +E2 13 hr (orange) using the Atp1a1 promoter as a viewpoint. Tracks are plotted on top of each other with the lowest one for each DpnII fragment depicted in grey. Significance of chromatin interactions was calculated using peakC (settings qWd:1.5, qWr:1.25, asterisk = significant interaction). F) Correlation between dREG cluster 7 and all gene clusters. Colors represent the odds ratio of observed over expected values before and after fitting a linear model with blue: lower than 1 and red: higher than 1. Number of asterisks indicates statistical significance of the enrichment based on a Fisher’s exact test: * P < 0.05, ** P < 0.01 and *** P < 0.001. Arrows indicate significant correlations between dREG cluster 7 peaks and downregulated gene clusters 7 and 11. Related to Figure S3 and Table S1.

We used the dREG algorithm to identify bidirectionally transcribed regions²⁷, representing putative regulatory elements, including TSSs and inter- as well as intragenic sites. Because intragenic dREG regions are confounded by reads on the sense strand, only antisense reads were counted in those cases (Figure S3A and Methods). A total of 26,991 dREG regions were found, and those with a significant pattern of change (ANOVA, P < 0.01) were grouped into 10 clusters. 6,011 (dREG clusters 1,2,5) and 1,271 (dREG cluster 8,9,10) putative regulatory regions (excluding TSSs) showed gradual signal losses (e.g. at the Myc enhancer) and gains (e.g. at the Mitoferrin-1 enhancer), respectively, albeit with varying kinetics (Figure 3B,C and Table S1). We detected 1,167 dREG peaks (dREG cluster 7) that appeared transiently, such as at the Prmt3 locus where, similar to the Kit gene, transient intragenic antisense transcripts arose at 13 hr of differentiation (Figure 3B,C). We examined the Atp1a1 gene as an additional example at which intergenic transient elements are found, in this case two GATA1 bound sites residing 41 kb and 52 kb downstream of the TSS (Figure 3D). 4C-seq using the promoter as a viewpoint showed increased proximity between the Atp1a1 promoter and the +41 kb site during repression (Figure 3E), providing an additional example of a newly established chromatin loop anchored by repression-specific regulatory sites.

To study the relationship between putative regulatory sites and target genes more globally, we coupled them based on a 500 kb window size. Accordingly, dREG peaks may be linked to multiple genes, but genes may also be coupled to multiple dREG sites (Figure S3B). We then identified combinations of regulatory elements and genes that occurred more frequently than expected (Figure 3F and Figure S3C; see Methods). Overall, a positive correlation was observed between the enrichment of gene cluster-dREG cluster combinations and the similarity of PRO-seq dynamics between gene expression and dREG peaks (Figure S3D). This suggests that, in general, putative regulatory regions and the genes close to them tend to be behave similarly over time. Therefore, we adjusted the expected frequency of all gene expression cluster-dREG cluster combinations by fitting a linear model. This revealed additional significant correlations (Fisher’s exact test) and, importantly, showed that transient dREG sites within cluster 7 are more frequently found at downregulated genes (Figure 3F). We interrogated how many gradually downregulated genes (clusters 2, 3, 7 and 11) have a cluster 7 peak within their 500 kb regulatory window. 1040 out of 1813 genes were linked to a cluster 7 dREG peak, suggesting these transiently active elements may delay silencing. Note that within this window size, a single dREG site can be linked to multiple genes not all of which have to be regulated by the element. The average distance to their linked promoters was 95 kb, and 45.93% and 54.07% were intragenic or intergenic, respectively. 1001/1040 (96.25%) of the repressed genes also contained initially active regulatory elements that were lost (dREG cluster 1, 2 or 5) following GATA1-ER activation. Therefore, as for Kit, loss of activating regulatory elements is likely a main driver of gene silencing. Importantly, over one thousand transient elements were identified, attesting to their widespread presence, and suggesting that repression kinetics are fine-tuned at numerous genes.

Normally transient regulatory elements persist upon failure to engage the FOG1/NuRD deacetylase complex

To investigate a potential direct involvement of GATA1 at dREG sites, we compared existing GATA1 ChIP-seq time series data sets (0, 3, 7, 14 and 24 hr of differentiation¹⁷) with our dREG clusters by direct overlap of peak lists (Figure 4A). In total, 40.5% of all putative dynamic dREG sites were GATA1-ER-occupied during at least one timepoint. This is consistent with the ability of GATA1-ER to decommission existing enhancers (e.g. cluster 1 and cluster 5) or create activating regulatory sites (e.g. cluster 10). Cluster 7 which contains transient dREG sites was the most enriched cluster for GATA1 binding with 68% of the regions being occupied by GATA1. The transient nature of cluster 7 regulatory sites suggested a switch from activation to self-antagonization during GATA1 occupancy.

Figure 4. Transient elements become persistent in the absence of FOG1/NuRD engagement.

A) Fraction of dREG peaks bound by GATA1-ER for each cluster. B) Anti-H3K27ac ChIP-seq tracks for G1E-ER4 (-E2, +E2 13 hr and +E2 24 hr) in green and G1E-GATA1(V205M)-ER (-E2, +E2 13 hr and +E2 24 hr) in black at the Kit locus. n=2 for each condition, average of replicates depicted. GATA1 bound regions of interest are highlighted by boxes. C) Kit primary transcript (pT) levels normalized to Gapdh and relative to the mean of -E2 samples for G1E-ER4 (n=3) and G1E-GATA1(V205M)-ER (n=3) cells. Error bars represent SEM. One-way ANOVA: ****P < 0.0001. D) Quantification of H3K27ac ChIP-seq enrichment at +49 kb transient element after GATA1-ER and GATA1(V205M)-ER activation for two independent replicates. Difference (calculated by subtraction) between rlog normalized read counts of indicated timepoints and -E2 are depicted. Error bars represent standard deviation (SD). E) H3K27ac dynamics at all GATA1-ER and GATA1(V205M)-ER bound cluster 7 dREG peaks in G1E-ER4 and G1E-GATA1(V205M)-ER cells. Relative enrichment represents number of standard deviations from G1E condition. Two-tailed, paired t-test: **** P < 0.0001. Related to Figure S4.

Friend of GATA (FOG1) is a GATA1 co-factor that binds the nucleosome remodeling and deacetylase complex NuRD^28–30, and GATA1 requires the FOG1/NuRD complex for Kit repression^28,31. In addition, the NuRD complex has previously been linked to the fine-tuning of gene expression during lineage commitment^32,33. We thus asked whether the eventual loss of H3K27ac at the Kit intronic elements is mediated by the FOG1/NuRD corepressor complex, which contains histone deacetylase activity, and whether failure to engage FOG1/NuRD might prevent the switch from active to inactive states. To this end we used a cell line expressing GATA1(V205M)-ER, an E2-dependent form of GATA1(V205M). GATA1(V205M) is a variant, originally discovered in patients with dyserythropoietic anemia, with reduced affinity for FOG1, impaired ability to silence Kit, and failure to induce normal erythroid maturation^34–36. GATA1(V205M)-ER occupied the −114 kb enhancer as well as the +5 and +49 kb intronic elements, as determined by anti-GATA1 ChIP-seq (Figure S4A). Even though present, GATA1(V205M)-ER activation failed to diminish H3K27ac at the −114 kb enhancer in accordance with sustained Kit expression (Figure 4B,C). At the +5 kb and +49 kb elements, GATA1(V205M)-ER augmented H3K27ac between 0 and 13 hr of E2 treatment, comparably to GATA1-ER. However, GATA1(V205M)-ER failed to reverse the H3K27ac signal but instead further increased it towards 24 hr (Figure 4B,D). Therefore, in the absence of FOG1/NuRD engagement, GATA1(V205M)-ER can normally develop the +5 kb and +49 kb sites into enhancer-like elements that may contribute to maintenance of Kit gene expression, but fails to nullify them. Similar regulatory dynamics were observed at the Atp1a1 locus (Figure S4B–F).

GATA1 binding was previously found to depend on FOG1 in a subset of the cases^37,38. In line with this, 372 (47%) of all GATA1-ER-bound dREG cluster 7 sites (n=792) were also found significantly enriched for GATA1(V205M)-ER. At these sites, GATA1-ER or GATA1(V205M)-ER binding resulted in an initial gain of H3K27ac up to 13 hr in 348 and 300 out of 372 putative regulatory elements, respectively. Notably, while followed by a H3K27ac reduction at 24 hr of GATA1-ER activation (on average 28.4%), GATA1(V205M)-ER activation further increased H3K27ac towards 24 hr (+14.2% on average, Figure 4E). Thus, failure to engage the FOG1/NuRD corepressor complex causes otherwise transient regulatory elements to become persistent.

Co-factor usage defines the output of GATA1-occupied sites

The action of NuRD at regulatory sites is complicated by the fact that distinct NuRD compositions and different modes of NuRD recruitment exist. Even though transcription factors can directly associate with NuRD^28,39–43, this only partially explains its genome-wide distribution. NuRD has also been described to, seemingly paradoxically, generally occupy open and active chromatin^{16,32,44–47}, and can even be required for gene activation³¹. To explore whether NuRD is regulated at the level of recruitment or whether its deacetylase activity is regulated at transient enhancers, we performed ChIP-seq of the two NuRD components CHD4 and MTA2. For each dREG cluster, GATA1-occupied sites were enriched for NuRD co-occupancy compared to sites without GATA1 binding (Figure S5A). At the +49 kb Kit element, CHD4 and MTA2 occupancy followed a pattern similar to that of H3K27ac and eRNAs with a transient increase between -E2 and 13 hr followed by a decrease towards 24 hr (Figure 5A,B and Figure S5B,C). Focusing on dREG cluster 7 elements bound by GATA1-ER and GATA1(V205M)-ER, NuRD dynamics appeared more complicated in that CHD4 and MTA2 did not display identical occupancy characteristics across time points (Figure S5D,E). This is likely a reflection of variable NuRD subunit compositions and distinct modes of regulation³⁹.

Figure 5. Co-factors define the output of GATA1-occupied sites.

A) Anti-CHD4 ChIP-seq tracks for -E2, +E2 13 hr and +E2 24 hr after GATA1-ER (G1E-ER4, red) as well as GATA1(V205M)-ER (G1E-GATA1(V205M)-ER, dark green) activation. n=2 for each condition, average of replicates depicted. B) Quantification of CHD4 ChIP-seq enrichment at +49 kb transient element depicted in (A) for G1E-ER4 and G1E-GATA1(V205M)-ER. Difference (calculated by subtraction) between rlog normalized read counts of indicated timepoints and -E2 are depicted. Error bars represent SD. C) Same as in (A) for LDB1. D) Same as in (B) for LDB1. Related to Figure S5.

In the context of GATA1(V205M)-ER, at the Kit locus CHD4 and MTA2 were recruited similarly between 0 and 13 hrs, however, both NuRD components persisted up to 24 hr (Figure 5A,B and Figure S5B,C). Therefore, NuRD occupancy mirrors active chromatin, in this case modulated by GATA1. This supports the notion that it is the engagement/tuning of NuRD activity rather than simply its recruitment that determines the function of the Kit regulatory sites.

Since co-factors seemed to define the output of GATA1-bound sites, we assessed occupancy of the well-known GATA1 co-activator LDB1, which is associated with GATA1-dependent gene activation and chromatin looping in erythroid cells^18,48–51. Globally, GATA1-bound dREG cluster 7 sites were enriched for LDB1 co-occupancy compared to elements without GATA1 (Figure S5A). In GATA1-ER expressing cells, 13 hr of E2 exposure triggered a modest increase in LDB1 occupancy at +49 kb compared to -E2 which was reduced at 24 hrs (Figure 5C,D). At the −114 kb and +5 kb elements, LDB1 was evicted at 13 hrs and further declined at 24 hrs. The GATA1(V205M)-ER mutant failed to evict LDB1 at −114 kb and +5 kb. In addition, while GATA1(V205M)-ER activation also triggered LDB1 occupancy at +49 kb, rather than a decrease towards 24 hr, LDB1 occupancy further increased (Figure 5C,D). A global analysis of GATA1-ER and GATA1(V205M)-ER bound transient dREG cluster 7 sites showed distinct LDB1 occupancy dynamics in the mutant as well (Figure S5F). Therefore, we speculate that as a result of reduced engagement of the FOG1/NuRD complex, GATA1(V205M)-ER forms a more persistent activating complex that, together with LDB1 and possibly other co-factors, perpetuates acetylation and activity of the intronic regulatory elements.

In sum, these findings support a model in which GATA1 triggers the formation of a transient enhancer element within the Kit intronic region that is ultimately limited by the FOG1/NuRD complex. The dynamics suggest that a regulatory site can quickly change identity based on co-factors even though its primary driver, in this case GATA1, remains the same. What underlies differential co-factor usage at distinct time points of differentiation remains to be determined.

Broad presence of transient elements across cell types and species

To assess whether the paradoxical appearance of enhancer marks during gene repression could be observed for Kit and other genes in primary erythroid cells, we started by investigating previously published data sets in primary fetal liver cells at distinct stages of erythroid maturation. In these studies, different populations were sorted based on CD71 and TER119 cell surface markers^52–54, with S0^lo cells representing hematopoietic stem cells and Burst-Forming Unit-Erythroid cells (BFU-Es; CD71^lo, TER119^lo), S0^med and S1 cells representing Colony-Forming Unit-Erythroid cells (CFU-Es; CD71^med to CD71^hi, TER119^lo). S2 (CD71^hi, TER119^med), S3 (CD71^hi, TER119^hi) and S4/S5 (CD71^lo,TER119^hi) represent more mature erythroid stages (Figure S6A). Analysis of the KIT (CD117) receptor protein in these populations using cite-seq antibody derived tags (cite-seq ADT⁵²) showed the surface protein is present in S0^lo, S0^med and S1 populations, with downregulation towards S2 and S3 (Figure S6B). Investigation of mature messenger RNA transcript counts (cite-seq RNA⁵²) confirmed Kit expression declined earlier than the protein (Figure S6C). Analysis of nascent transcripts⁵⁴ revealed that transcriptional silencing of the erythroid precursor-specific genes Kit, Myc and Gata2 starts between S0^lo, S0^med and S1, suggesting that repression is initiated in the earliest, more rare, BFU-E and early CFU-E populations (Figure S6C–E). GATA1 transcripts increased slightly between S0^lo and S0^med (Figure S6F), but not to a similar extent as the sharp decrease of mostly Gata2 expression between S0^lo and S0^med. This is consistent with previous observations that GATA1 protein, not mRNA, increases rapidly in early stages of human erythroid differentiation⁵⁵.

Because of limited cell numbers of the progenitor cells, we performed H3K27ac CUT&RUN on 100,000 cells for each population directly after sorting primary cells from fetal livers based on CD71 and TER119 (see Methods). S0^lo and S0^med cells express the early marker gene Ptprc (also known as CD45), which is enriched for H3K27ac in these populations, while in S1 cells this gene is silenced (Figure S6G). Similar findings were made for CD34 (Figure S6H). This confirms the S0 cells are myeloid progenitors and BFU-Es. Accordingly, H3K27ac at the CD71 (Tfrc gene) locus was lowly enriched in S0 cells, while gaining H3K27ac in S1 cells and later stages (Figure S6I). Unfortunately, CUT&RUN signal at TSSs in general as well as intronic sites within the Kit locus were too sparse to draw firm conclusions (Figure S6J). However, at the Atp1a1 gene which is downregulated later than the Kit gene, transient H3K27ac enrichment was observed at the +52 kb and +41 kb sites between S1 and S4 (Figure S6K,L), reminiscent of our findings in G1E-ER4 cells (Figure 3D and Figure S4B).

Since H3K27ac CUT&RUN signals were not strong enough at several regulatory regions of interest, we carried out ChIP-seq. To obtain ~5 million cells per population, S0 cells were purified from E13.5 fetal liver cells and analyzed immediately (KIT^hi, CD71^lo) or cultured in medium containing stem cell factor, erythropoietin and dexamethasone for one day (KIT^hi, CD71^med) as described previously^56,57 (Figure 6A). For more differentiated populations, cells were expanded longer and sorted using FACS based on KIT and CD71 levels to collect erythroid cells at finely spaced maturation stages (KIT^hi CD71^hi to KIT^lo CD71^hi, six populations with KIT levels from high (100) to low (0); and KIT^lo CD71^med; Figure 6A). Overall, H3K27ac enrichment at regulatory (excluding TSSs) sites was highly correlated with the CUT&RUN data (Figure S7A). Likewise, genes indicative of precursor stages, Gata2 and Myc, as well as those reflecting terminal maturation Hbb1 and Gypa (encoding TER119) along H3K27ac enrichment at the Ptprc, Cd34 and Tfrc loci demonstrated the usefulness of the cell sorting strategy (Figure 6D and Figure S7B–E).

Figure 6. Transient regulatory elements in primary murine cells during erythroid differentiation.

A) Strategy to isolate and analyze primary mouse fetal liver cells at different stages of differentiation based on cell surface markers KIT (CD117) and CD71. B) Anti-H3K27ac ChIP-seq tracks at the Atp1a1 locus for populations from mouse fetal liver cells described in (A). C) Same as in (B) for the Kit locus. D) RT-qPCR for Kit primary transcripts (pT) in populations described in (A). E) Same as in (B) for the Eng locus. F) Same as in (B) for the Pex14 locus. G) Same as in (B) for the Mrpl23 locus. Related to Figure S6 and S7.

59% of the GATA1-bound cluster 7 dREG peaks with H3K27ac enrichment in G1E-ER4 cells were also observed in mouse fetal livers cells. The difference seemed to be due to lower peaks not being called in primary cells perhaps as a result of lack of synchrony (Figure S7F). Analysis of shared sites revealed transient H3K27ac dynamics in primary cells as well (Figure S7G). The high level of stratification of the cells enabled further classification of H3K27ac peaks and 49% of the elements peaked in KIT^hi to KIT^lo(20) stages, including the Atp1a1 +41 kb element (Figure 6B and Figure S7H). The different kinetics for the Atp1a1 +41 kb and +52 kb elements further suggested that self-limitation depends on distinct co-factors or distinct concentrations of the same co-factors (Figure 6B and Figure S7H). H3K27ac enrichment Kit +5 kb and +49 kb regions peaked in KIT^hiCD71^hi populations (Figure 6C and Figure S7H) in which Kit gene transcription is being downregulated (Figure 6D). Our sorted populations were, however, not homogenous enough to reveal the concomitant loss of the −114 kb enhancer and gain of the intronic elements observed in highly synchronized G1E-ER4 cells upon GATA1-ER activation.

Nonetheless, since 59% of the transient regulatory sites identified in G1E-ER4 cells were also found in murine fetal liver cells (Figure S7F,G), this confirms the presence of transient elements in primary cells during differentiation. Additional examples of downregulated genes with nearby transient elements are provided by the Eng, Pex14 and Mrpl23 loci (Figure 6E–G and Figure S7H,I).

To extend our analysis to a different cell type and species, we used published enhancer activity profiles during differentiation stages from mesenchymal stem cells to adipocytes⁵⁸. Gene clusters (based on RNA-seq) and enhancer clusters (based on Mediator 1 (Med1) ChIP-seq) described in Rauch et al.⁵⁸ were analyzed (Figure S7J). Adipocyte enhancer cluster 4 (n = 1,382) showed similar behavior as our dREG7 sites with a sharp increase towards 4 hr after induction of differentiation and loss of enrichment towards 24 hr and later time points. While clusters were defined based on Med1 dynamics, H3K27ac at the same sites also showed a transient pattern (Figure S7K). We calculated correlations using the same parameters as used for our data in Figure 3F and Figure S3B,C and found a similar genome-wide enrichment between transient putative enhancer elements and downregulated genes (Figure S7L). The MTMR2 locus is shown as an example, which comparable to Kit in murine G1E-ER4 cells, has an intronic transient element that arises during transcriptional downregulation, while another enhancer is gradually inactivated (Figure S7M,N). This suggests that transient putative enhancer-like elements modulate gene repression dynamics across cellular lineages and species.

DISCUSSION

To dissect mechanisms of transcriptional repression, we profiled histone marks, nascent transcription, and chromatin looping at various time points during Kit repression by GATA1. Concomitantly with the decommissioning and “un-looping” of a major upstream Kit enhancer, GATA1 triggers the formation of transient regulatory elements with enhancer-like features (H3K27ac, eRNAs, long-range chromatin contacts) that delay Kit silencing. Moreover, the transient enhancer is capable of maintaining a degree of Kit gene activity in cells lacking the −114 kb enhancer. GATA1 is therefore not only a transcription factor with the ability to activate and repress distinct gene sets, but, as revealed here, can exert the two roles at a single locus and even at single elements. These findings point towards a novel mechanism through which a transcription factor self-modulates its gene silencing activity.

What causes the rapid transition from active to inactive at the transient elements? It is possible that GATA1 initially promotes histone acetylation by cooperating with CREB-binding protein and p300⁵⁹. These coactivators can also acetylate GATA1⁶⁰, thereby temporarily stabilizing it on chromatin, perhaps via acetyl histone readers of the BET family, which in turn can bind both acetylated histones and transcription factors via their bromodomains ^61,62. This might expose the substrate for subsequent action by the NuRD complex, as NuRD engagement can also be mediated by BET proteins⁶³ or direct acetyl lysine binding via the CHD4 PHD domain⁶⁴. At this point it remains challenging to discern to what extent and in which order deacetylation of histones and possibly of GATA1 limit the action of this element.

Some regulatory sequences that repress transcription in one cell type can function as enhancers in others, potentially because they are bound by distinct transcription factors^65–70. However, in the present system, functionality is mediated by the same transcription factor i.e. GATA1, but the output is determined by co-factor dynamics at the same site. This is suggested by our observation that the disease causing mutant GATA1(V205M)-ER with reduced affinity for FOG1 led to persistence of otherwise transient elements. An example for contextual co-factor modulation was provided by earlier studies showing that specific ETS factors, when positioned near GATA1 bound sites, can determine whether FOG1 functions as a GATA1 co-activator or co-repressor^71,72. Since the NuRD components MTA2 and CHD4 are present at putative transient enhancers in both GATA1-ER and GATA1(V205M)-ER expressing cells, additional proteins, including but likely not limited to LDB1, may drive activation or inhibition of NuRD activity.

The presence of multiple GATA1 binding sites at the Kit locus (e.g. the +33 kb transient element) suggests that gene silencing kinetics might be further modulated by cooperative or consecutive action of different elements. An open question remains whether upon NuRD engagement (between 13 and 24 hr), the +49 kb and +5 kb elements simply lose their activity, or whether they actively contribute to transcriptional repression of the Kit gene. The removal of acetyl marks on histones or transcription factors by NuRD might argue in favor of an active repression mechanism. A series of very closely spaced time points at a single allele level will be needed to fully appreciate the dynamics and (potentially opposite) functional consequences of the different cis-regulatory elements. Ultimately, the repressed state is maintained by heterochromatin formation in the absence of binding of the silencing-initiating factor GATA1.

Albeit to a distinct degree, both the −114 kb enhancer and +49 kb transient element are able to maintain transcription of the Kit gene. The +49 kb element is reminiscent of shadow enhancers, which may be redundant in certain contexts but can buffer against transcriptional changes, for example, when a different enhancer of the same gene is lost, or under cellular stress conditions^73–77. However, the transient enhancer at the Kit gene seems distinct from shadow enhancers in several ways. First, the transient enhancer is not redundant because its deletion accelerates Kit silencing. Its activity peaks at time points during differentiation that are different from the −114 kb enhancer, which also argues against redundancy. Second, transient enhancers are short-lived and self-limiting. Notably, the same transcription factor, GATA1, initiates loss of the −114 kb enhancer while concomitantly activating the transient enhancer. Third, numerous transient enhancers in erythroid and non-erythroid cells appear specifically during silencing of nearby genes. To our knowledge, shadow enhancers have not been described in the context of gene repression but have been considered in other contexts, such as spatial patterning or stress conditions.

Different transient putative enhancers peaked at distinct stages of erythroid differentiation in primary cells, which indicates that sampling at high temporal resolution is essential to identify them. While their function remains to be tested via perturbative experiments in more tissues, our data suggest that transient enhancers may optimize gene repression kinetics. In the case of the Kit gene, we speculate that the transient enhancer prevents too rapid of a Kit shut down because it might be detrimental to viability of the cells as they await the activation of other pro-survivals genes, such as the GATA1 target gene Bcl-xL⁷⁸. Moreover, transient enhancers may offer an avenue to extend the proliferative phase without blocking the initiation of cell maturation. Manipulating the balance between proliferation and differentiation of erythroid precursors could aid ex vivo red blood cell production, a process which is being explored to meet the high clinical demand for red blood cells. The in-depth investigation of distinct steps of gene silencing, and a possible role for elements that oppose activating signals or delay repression, will help to better understand cell fate decisions and could be translated into applications that target the mechanisms and/or timing of gene silencing.

LIMITATIONS OF THE STUDY

Our study used an immortalized cell line that enabled highly synchronous control of Kit gene repression and gene editing, which led to the identification of elements with unexpected properties. While we also found candidate transient enhancer elements in primary differentiating erythroid cells, testing them will require their conditional deletion in whole animals, followed by analysis of highly pure erythroid populations at finely spaced stages of maturation.

STAR Methods text

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by Gerd A. Blobel (blobel@chop.edu)

Materials availability

Unique/stable reagents or cell lines generated in this study are available upon request to the lead contact.

Data and code availability

• Datasets reported in this study are available at the Gene Expression Omnibus as part of SuperSeries GSE193825 with accession numbers: GSE193718 (4C-seq G1E-ER4), GSE193740 (ChIP-seq G1E-ER4), GSE223244 (ChIP-seq mouse fetal liver), GSE193741 (PRO-seq G1E-ER4) and GSE171385 (CUT&RUN mouse fetal liver).

• The code used to analyze our PRO-seq data can be accessed through: https://github.com/zhezhangsh/PROseqR and zenodo DOI: 10.5281/zenodo.7601361.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

The G1E erythroblast cell line was derived from GATA1−/− mouse embryonic stem cells¹⁴. The subline G1E-ER4 expresses GATA1 fused to the ligand binding domain of the estrogen receptor. Addition of 100 nM estradiol to the growth medium activates GATA1 and induces erythroid maturation. To obtain Kit cDNA overexpressing cells (G1E-ER4k cells), the murine Kit cDNA sequence and YFP were cloned into a MigR1 retroviral vector. G1E-ER4 cells were transduced and YFP+ cells were sorted using a Beckman Coulter Moflo Astrios sorter.

METHODS DETAILS

Cell culture

G1E, G1E-ER4 and G1E-GATA1(V205M)-ER cells were grown in IMDM media (Corning, 10–016-CV) with 15% fetal calf serum (Gemini, lot A08H00K), 2% penicillin/streptomycin (Gibco, 15140–122), 140 μm 1-thioglycerol (Sigma, M6145), 2 units/ml epoetin (Amgen) and conditioned medium from a kit ligand-producing CHO cell line. Cells were grown at a density of 100,000 to 1,000,000 cells/ml. Erythroid maturation experiments were performed by plating cells at a cell density of ~200,000 cells/ml in fresh G1E medium containing 100 nm estradiol (Sigma, E2758) for durations indicated for each experiment.

Retroviral transduction of G1E-ER4 cells

MigR1-Kit cDNA-YFP retrovirus was produced by co-transfecting 10 μg of construct and 10 μg pCL-Eco packaging plasmid into 293T cells using polyethylenimine (PEI, Polysciences Cat#23966). Virus-containing medium was harvested after 24 and 48 hr and passed through a 0.2 μm filter. Retroviral transduction was performed by adding 500 ml of viral supernatant supplemented with 4 μl/ml Polybrene and 10 μl/ml 1 m HEPES (Gibco, 15630080) to between 200,000 and 500,000 cells in 500 ul G1E-ER4 medium in 12-well plates by spinfection at 2300 g for 90 min. Supernatant was removed and cells were resuspended in fresh medium after spinfection. YFP positive cells were sorted 24 hr after transduction.

CRISPR/Cas9-mediated genome editing

12 μg of gRNA and Cas9-containing plasmid (PX330 mCherry, Addgene Plasmid #98750) was transfected into 3 million cells using the Amaxa II electroporator (Lonza; program G-016) and Amaxa II Cell Line Nucleofector Kit (R) (Lonza, VCA-1001). An empty PX330 vector transfected sample was used as control (“unedited”). mCherry positive cells were sorted 24 hr after nucleofection using a Beckman Coulter Moflo Astrios sorter. Single-cell clones were expanded for 6 days, followed by colony screening using standard PCR. Unedited control clones went through the procedure of nucleoporation and sorting. Guide sequences to target Kit +49 kb element (190 bp deletion): gRNA#1 GACTCTAAGCGGGTTGTAGA and gRNA#2 CAGCTAGTAAAGCCCTACTC; Kit −114 kb GATA motif single gRNA AGATTATCTCAAGACAACAC, Kit −114 kb (100 bp deletion) gRNA#1 TCTGAGATGCGGTTGCTGAT and gRNA#2 TCCCATGACTCTAACGTGCC, Kit −114 kb (200 bp deletion) gRNA#1 CCGCTGAGTTGCTAAGGGCC and gRNA#2 TCTGAGATGCGGTTGCTGAT. For double deletion of both Δ+49 kb and Δ−114 kb elements, G1E-ER4k cells were first transfected with Δ+49 guides. mCherry positive cells were sorted and cultured until there were 3 million cells for a second round of transfection with Δ−114 kb guides. Single cells clones were sorted and screened for both edits. The following primers were used to screen for homozygously deleted clones: Δ+49 deletion CAGTGGGAGCATACAAAACTC and CCTGGTGGACCGAATGAG; Δ−114 kb (100 bp and 200 bp deletion) CAGGAATTGATGAAATGCTGACC and CATCTTCTGCTTTTGAGCTTAGAAC. Genotypes of selected clones were confirmed by Sanger sequencing.

RNA isolation and RT-qPCR

Desired number of cells were pelleted and resuspended in 400 μl RLT buffer. RNA was purified according to manufacturer’s protocol. Columns were treated with Rnase free DNase (Qiagen, 74106) on column and eluted in RNase free water. Concentration and purity were measured using a spectrophotometer (Nanodrop2000, Thermoscientific) and cDNA was made using a High-Capacity cDNA Reverse Transcription Kit (Thermofisher scientific, 4368814). RT-qPCR was performed using 5–20 ng of cDNA in 4 μl, 1 μl of 10 μm forward/reverse primers, and 5 μl of Power SYBR^™ Green PCR Master Mix (Thermo Fisher Scientific, 4367660) on an ABI Vii7 real-time PCR machine. Primer sequences for RT–qPCR: Kit mature transcript AGCAGATCTCGGACAGCACC and TGCAGTTTGCCAAGTTGGAG; Kit primary transcript AACTGAAGCGAGTACAGCATTCC and TGCTTTTGCTTGTGTACT GTTAACTG; Atp1a1 primary transcript CTTTAAGAGCGCCGACTCAGAGG and GCCTTTCCAGGTCACTAACAGACGATG. Gata2 primary transcript AGAGCCCATGGTCTAGCAGC and CGGCCAGAGCAGCCATAG; Myc primary transcript TAGACTTGCTTCCCTTGCTG and TCGTCGCAGATGAAATAGGG; Hbb-b1 primary transcript GCCTGCAGTATCTGGTATTTTTG and TGAAATCCTTGCCCAGGTG; Gypa primary transcript CTGTGGGCTGTGGGAAGCATTTTATC and CCACCACAGGAGTGGTTATTTGAAGAAG.

Chromatin Immuno Precipitation (ChIP)

Cells were crosslinked in 1% formaldehyde in PBS/2mM EDTA at room temperature for 10 minutes. Glycine was added to a final concentration of 1M and samples were rocked for another 5 minutes. Cells were pelleted by centrifugation and washed twice with cold PBS. Pellets of 20 million cells were flash frozen and thawed on ice on the day of starting the actual ChIP experiment. Cell pellets were resuspended in 1 ml cell lysis buffer (10mM Tris pH 8, 0.2% Igepal, 10 mM NaCl) with protease inhibitors (Sigma P8340) and 1mM phenylmethylsulfonyl fluoride (PMSF) and incubated on ice for 20 minutes. Nuclei were pelleted and resuspended in 1 ml nuclear lysis buffer (50 mM Tris pH 8, 1% SDS, 10 mM EDTA with protease inhibitors and PMSF), and incubated on ice for 20 minutes. Samples were then sonicated for 45 minutes at 4°C using the following settings 100% amplitude, 30 sec on/30 sec off (Epishear, Active Motif). After sonication, samples were spun at 21130×g for 5 minutes at 4°C. Supernatant (~ 1 ml) was diluted using 4 ml IP Dilution Buffer (20 mM Tris pH 8.0, 1 M Tris pH 8.0, 2 mM EDTA, 150 mM NaCl, 1% Triton X-100, 0.01% SDS with protease inhibitors). Samples were precleared using 50 μg of isotope-matched IgG and 50 ul A/G agarose beads (prepared by mixing Protein A (ThermoFisher 15918014) and Protein G (ThermoFisher 15920010) at 1:1 ratio) for ≥2 hr. 200 μl of precleared chromatin was set aside as input and 2.2 ml was used to set up two IP reactions by adding A/G beads pre-bound with antibody of interest and rotated overnight at 4°C. On the second day, beads were washed once with IP Wash 1 (20mM Tris pH 8, 2mM EDTA, 50 mM NaCl, 1% Triton X-100, 0.1% SDS), twice with High Salt Buffer (20 mM Tris pH 8, 2mM EDTA, 500 mM NaCl, 1% Triton X-100, 0.01% SDS), once with IP Wash Buffer 2 (10 mM Tris pH 8, 1mM EDTA, 0.25 M LiCI, 1% Igepal, 1% sodium deoxycholate), and twice with TE (10mM Tris pH 8, 1mM EDTA pH 8). All washes were performed on ice after which beads were eluted twice with 100 ul Elution Buffer (100mM NaHCO3, 1% SDS, prepared fresh). 12ul of 5M NaCl, 2 μl RNaseA (10mg/ml, 10109169001 BMB) were added to samples as well as inputs and put at 37°C for 30 minutes. 3 μl of proteinase K (20mg/ml, 3115879 BMB) was added and samples were incubated at 65C overnight. Next, 10 μl of 3M sodium acetate pH 5.2 was added to each sample and DNA was purified using the QIAquick PCR Purification kit (Qiagen 28106) per the manufacturer’s instructions. The following antibodies were used: anti-H3K27ac (Active Motif, #39133), antiGATA1 (Santa Cruz Biotechnology, #SC-265), anti-H3K27me3 (Cell Signaling, #9733S) ) and anti-LDB1 (Santa Cruz, sc-365074). For experiments using antibodies against CHD4 (Abcam, #ab70469) and MTA2 (Abcam, #ab8106), the same ChIP protocol was used with one adaptation. Cells were crosslinked in PBS/2mM EDTA with 1.5mM EGS for 20 minutes at room temperature, followed by addition of formaldehyde to a final concentration of 1% with crosslinking for an additional 10 minutes at room temperature.

ChIP-qPCR

Quantitative PCR after ChIP was performed using Power SYBR^™ Green PCR Master Mix (Thermo Fisher Scientific, 4367660) on an ABI Vii7 real-time PCR machine. The following primers were used: Kit promoter GCGTGGCAGCAACGAAC and AGCCGCTAGGTTGCAGACTTT; −114 kb Kit enhancer GGGCAGGCACTAAATTCAAATCAGTAAG and GCTTCAGGTTGGTCTTCCTTGG; +5 kb Kit GTAGACACCCGCACAGGAG and CAGATGTTCGCTCCCCTTCTG; +49 kb Kit GCTGTGAGAGAGTGCACAAGC and CCATGCCCTCGCTGTTAAGATGG.

ChIP-Sequencing

ChIP-seq libraries were prepared using Illumina’s TruSeq ChIP sample preparation kit (Illumina, Cat#IP-202–1012) according to manufacturer’s protocol. Size selection was performed using SPRIselect beads (Beckman Coulter, Cat#B23318) (left side at 0.9x, right side at 0.6x). Libraries were sequenced (1×75bp) on the Illumina NextSeq 500 or 2000 platform according to manufacturer’s instructions. Bclfastq2 v 2.15.04 (default parameters) was used to convert reads to fastq.

Precision Run On (PRO)-Sequencing

PRO-seq experiments were performed as previously reported⁷⁹. Two PRO-seq replicates were prepared for G1E as well as G1E-ER4 cells at 13 and 24 hr of differentiation. Three replicates were sequenced for G1E-ER4 before differentiation (-E2). For undifferentiated (-E2) and G1E-ER4 +E2 13 hr, 50 million cells were used. For G1E-ER4 +E2 24 hr experiments, 100 million G1E-ER4 cells were used. Size-selected libraries were pooled and sequenced (2×75bp) on the Illumina NextSeq 500 or 2500 platform.

4C Sequencing

To generate 4C template, 10 million cells were fixed and processed as previously described in detail⁸⁰. DpnII and Csp6I were used as first and second restriction enzyme, respectively. The following 4C primers were used to amplify viewpoints of interest (italic nucleotides are adapter sequences) using 150–200 ng 4C template input in 4 independent PCR reactions (4 × 50 μl) for each viewpoint: Kit promoter TACACGACGCTCTTCCGATCTGGTTCTTTAAGTCCCAAGATTGTA and ACTGGAGTTCAGACGTGTGCTCTTCCGATCGACAGCCAGTAACCCTGTC ; Kit +5 kb TACACGACGCTCTTCCGATCTCAAAGACTAGATGTTCCTCAAGATC and ACTGGAGTTCAGACGTGTGCTCTTCCGATCAGGTCAAGATGAAGGCAGAG; Kit +49 kb TACACGACGCTCTTCCGATCTGACCCAGCTCAATGGATC and ACTGGAGTTCAGACGTGTGCTCTTCCGATCCTCAGAAATCCACAGCACTT; Atp1a1 promoter TACACGACGCTCTTCCGATCTCACGATGGAAACCGGATC and ACTGGAGTTCAGACGTGTGCTCTTCCGATCGTGAGACAGCAGTGCAGGA. Reactions were pooled and ¼ of the total volume (200 μl) was purified using Ampure XP beads (Beckman Coulter, #A63880) eluting in 35 μl milliQ water. 10 μl purified PCR product was used as input material for a second PCR to obtain full Illumina adapter sequences using a universal forward primer (AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACGCTCTTCCGATCT) and a barcoded reverse primer (CAAGCAGAAGACGGCATACGAGATxrefGTGACTGGAGTTCAGACGTGTGCT) with a 6 nucleotide barcode (xref). PCR reactions were purified using a Roche High Pure PCR purification kit (#11732676001). Pooled 4C libraries were purified with Ampure XP beads before sequencing on the Illumina NextSeq 500 platform (1×75 bp).

Mouse fetal liver isolation with sorting based on CD71 and TER119 followed by H3K27ac CUT&RUN

Primary erythroid progenitor cells were isolated from fresh fetal livers from C57BL/6 mouse embryos at E12.5-E13.5 as previously described . For each experimental replicate, 5–15 livers were pooled and mechanically dissociated in staining buffer (PBS, 0.2% BSA, 5 mM glucose) followed by straining through a 30 micron strainer. To block Fc receptors, cells were immunostained at 4°C with rabbit IgG (200 μg/ml, Jackson Laboratories 015–000-003). Early erythroid progenitors were enriched by staining cells with 5 μg/ml biotin-conjugated anti-Ter119 (BD 553672) for 30 min followed by magnetic depletion using streptavidin nanobeads (BioLegend Mojosort 480016) according to the manufacturer’s instructions. Then, cells were incubated with a panel of 5 FITC-conjugated lineage antibodies (anti-CD41, anti-CD45R, anti-CD3e, anti-CD11b and anti-Ly-6G/6C, all at 1 ug/ml; BD 553848, 553087, 553061, 557396, and 553126, respectively) as well as 0.5 μg/ml APC-conjugated streptavidin (BD 553672) and 0.33 μg/ml PE-Cy7-conjugated anti-CD71 (BioLegend 113811) for 45 minutes. 0.66 μg/ml Hoechst was added immediately prior to sorting in FACS running buffer (staining buffer plus 2 mM EDTA). Live cells were sorted into Eppendorf tubes containing 500 ul RPMI supplemented with 10% FCS using a BD FACSAria Fusion machine (nozzle size, 100 μM).

CUT&RUN was carried out as described in⁸¹. Briefly, binding buffer (20 mM HEPES-KOH pH 7.9, 10 mM KCl, 1 mM CaCl2, 1 mM MnCl2) was used to wash Concanavalin A magnetic beads twice. 100,000 FACS-sorted erythroblasts were washed twice in 1.5 ml of wash buffer (20 mM HEPES pH 7.5, 150 mM NaCl, 0.5 mM spermidine and one Roche Complete Protease Inhibitor tablet) followed by resuspension in 200 μl wash buffer with 10 μl activated bead suspension. Samples were rotated for 5 to 10 minutes at room temperature. Tubes were placed in a magnetic rack, wash buffer was removed and replaced with 200 μl of antibody buffer (wash buffer plus 0.02% digitonin, 2 mM EDTA and 1 μl anti-H3K27ac Rabbit monoclonal antibody (Millipore; MABE647)). Samples were rotated for 2 hours at 4°C followed by one wash with digitonin buffer (wash buffer plus 0.02% digitonin) before resuspension in 200 μl digitonin buffer. pA-MNase enzyme, a kind gift of the Henikoff lab, was added to each sample to a final concentration of 700 ng/ml and samples were rotated for 1 hour at 4 C. Cells were washed twice with digitonin buffer and resuspended in 150 μl digitonin buffer. Digestion was initiated by the addition of 3 μl 100 mM CaCl2 and proceeded whilst tubes incubated on ice for 30 minutes. Reactions were quenched by addition of 50 μl 4x stop buffer (680 mM NaCl, 40 mM EDTA, 8 mM EGTA, 0.04% digitonin, 0.1 mg/ml of RNase A, 0.1 mg/ml glycogen), incubation for 10 minutes at 37 C and centrifugation for 5 minutes at 4 C and 16,000xg. Supernatant was moved to fresh tubes and incubated for 10 min at 70 C with 1 μl 20% SDS and 1.5 μl 20 mg/ml proteinase K. DNA was purified using phenol-chloroform extraction and ethanol precipitation. Libraries were prepared using the Ultra-II DNA library prep kit (NEB) with 13 PCR cycles according to the manufacturer’s protocol.

Mouse fetal liver isolation with sorting based on KIT(CD117) and CD71 followed by H3K27ac ChIP-seq

All mice were bread and cared for in accordance with CHOP IACUC-approved protocols. Primary erythroid progenitor cells were isolated from fresh fetal livers from C57BL/6 mouse embryos at e13.5 and expanded as described in⁵⁶. Briefly, fetal livers were made into a single cell suspension by pipetting up and down in PBS with 2% FBS, 2.5mM EDTA and 10mM glucose using a P1000 tip. Cells were passed through a 40 micron strainer into a FACS tube and spun down for 3 minutes at 1200 RPM at 4°C. Cells were resuspended in PBS with 2% FBS, 2.5mM EDTA and 10mM glucose and hematopoietic progenitors were purified using the Stemcell EasySep Mouse Hematopoietic Progenitor Cell Isolation Kit (Cat 19856) with 5 μL/mL CD71-biotin (EBioscience) according to manufacturer’s protocol. Isolated progenitors were expanded in StemPro-34 (ThermoFisher) media with 1× nutrient supplement, 2 mM L-Glutamine, 1% penicillin/streptomycin, 100 μM monothioglycerol, 1 μM dexamethasone, 0.5 U/ ml of erythropoietin (Epo), and 1% conditioned medium from a stem cell factor (SCF) producing CHO cell line. The following populations were analyzed: 1) progenitors immediately after isolation (KIT^hi, CD71^lo, no sort to have high enough cell numbers for ChIP, ~5 million). 2) progenitors expanded for one day in the culture media described above (KIT^hi CD71^med, no sort to have high enough cell numbers for ChIP). 3) six populations between KIT^hi CD71^hi to KIT^lo CD71^hi sorted based on KIT cell membrane levels after 2 days of expansion. 4) KIT^lo CD71^med cells sorted after 4 days of expansion. Since progenitors expanded in the described medium are not expected to differentiate further than erythroblasts, we did not try to sort more mature populations. Prior to sorting, cells were crosslinked in PBS/2mM EDTA with 1% final concentration formaldehyde for 10 minutes at room temperature. Glycine was added to a final concentration of 1M and samples were rocked for another 5 minutes. Cells were pelleted by centrifugation and washed twice with cold PBS/2mM EDTA. Fixed cells were transferred to an Eppendorf tube that was precoated with PBS/BSA to prevent the cells from sticking to the side of the tube and stained for 15 minutes at room temperature with 4 μg/ml KIT(CD117)-APC (Biolegend 105812) and 0.4 μg/ml CD71-PE (Biolegend 113808) in staining buffer with protease inhibitors (Sigma P8340) and 1mM PMSF. Cells were washed in 1 ml PBS, spun down and resuspended in 1x FACS buffer with PI (2μl/mL) and PMSF (10μl/mL) for sorting on a Cytek Aurora machine (nozzle size, 100 μM).

QUANTIFICATION AND STATISTICAL ANALYSIS

ChIP-Sequencing and CUT&RUN data processing and analysis

Sequences were aligned to mm9 using Bowtie 1.1.0⁸². Reads with more than one mismatch or multiple alignments were excluded. MACS2 version 2.1.0⁸³ was used to call significantly enriched regions (parameters: p = 10^−5, extsize = 300 and local lambda = 100,000) using whole-cell extract input controls. Bigwigs were RPM normalized. To obtain a list of all H3K27ac enriched regions across the 4 timepoints in G1E-ER4 cells, peaks were called for each sample separately and combined by merging overlapping peaks to obtain one reference list of 29,448 H3K27ac-enriched sites. Read coverage was determined with bedtools⁸⁴ and rlog normalized using DESeq2⁸⁵. The same strategy was applied for H3K27ac of mouse fetal liver CUT&RUN as well as ChIP-seq experiments for the distinct stages of differentiation assessed. When depicting rlog fold change for one regulatory element (Figure 4D, 5B, 5D and Figures S1B, S4D, S4E, S5C and S7H), control rlog values (either G1E or -E2) were subtracted from the rlog value of more differentiated conditions. When a list of elements was analyzed (Figures 4E and S5D–F), rlog values were further normalized by calculating the number of standard errors between the (differentiated) sample and control condition (G1E or -E2).

PRO-Sequencing data processing and analysis

PRO-seq data were processed as described previously⁷⁹, the code for which can be accessed through https://github.com/zhezhangsh/PROseqR. In short, read pairs originating from short RNA fragments were merged with FLASH2⁸⁶. PCR duplicates were removed using 6-base UMI barcodes, which were trimmed from the reads afterwards. Both merged single end reads and unmerged paired end reads were aligned to reference genome (mm9) using BWA-MEM⁸⁷. Primary alignments with no INDELs and mapq score of at least 20 were reported.

To identify putative cis-regulatory elements, bidirectionally transcribed peaks were called based on 3’ nucleotide alignment positions using the dREG algorithm (https://github.com/Danko-Lab/dREG). Dynamics across timepoints were compared between the two strands of each dREG peak, and split into two half peaks if the strands were not correlated. If dREG peaks had correlated dynamics between the two strands, they remained as one full peak. dREG peaks at least 1 kb away from any known TSSs were categorized into three types of regulatory elements. 13,726 full/half dREG peaks were intergenic, having no overlap to any known genes. 6,283 half dREG peaks overlapped the antisense strand of annotated genes. 8,896 full/half dREG peaks overlapped the sense strand of annotated. Half sense peaks were not analyzed as they are confounded by reads that result from expression of the gene. Full sense peaks were further curated and only used for analysis when the full dREG peak had significantly higher sequencing coverage than its surrounding intragenic areas. Numbers of merged single end reads or unmerged read pairs aligned to full/half dREG peaks including TSSs (n=13,173) as well as known gene bodies (to analyze gene expression) in all PRO-seq libraries were tallied to generate read count matrixes that are listed in Table S1.

Clustering analysis of dREG peaks and known genes was performed after read counts were normalized, log₂-transformed, further adjusted to the average of G1E libraries and rescaled to make the standard deviation of each gene or dREG peak equal to 1.0. Genes and peaks with significant changes across time points (ANOVA P < 0.01) were clustered based on their correlation to each other. Clusters were defined by similar genes or peaks within the same branches on a hierarchical tree, and further optimized to minimize their within-cluster variance and maximize their between-cluster variance.

We used a 500 kb window to map dREG peaks to known genes as their regulators (Figure S3B). The observed frequency of each combination between peak and gene clusters was compared to its expected frequency using Fisher’s exact test. As transcribed loci near each other tend to have simultaneous changes across timepoints, we postulated that peaks and genes mapped to each other were more likely to be grouped into the clusters having similar dynamics (Figure 3F and Figure S3C). Indeed, the observed frequencies of cluster-cluster combinations have a linear correlation to the similarity between peak-gene clusters (Figure S3D). We fitted the observed frequencies and cluster similarity to a linear regression model, and then replaced the expected frequencies of cluster combinations with the values predicted by the linear model. The adjusted expected frequency of each combination were used as the hypothesized odds ratio in Fisher’s exact test, instead of the default value of 1.0.

4C-Sequencing data processing and analysis

4C mapping was done as described in https://github.com/deWitLab/4C_mapping. PeakC was used to call peaks that are significantly higher (settings qWd:1.5, qWr:1.25) than background, which is modelled using monotonic regression²⁴ (https://github.com/deWitLab/peakC).

Supplementary Material

3

Table S1. Raw PRO-seq counts within gene bodies and dREG peaks at different time points of differentiation. Related to Figure 3 and Figure S3.

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Anti-H3K27ac	Active Motif	Cat#39133
Anti-GATA1	Santa Cruz Biotechnology	Cat#SC-265
Anti-H3K27me3	Cell Signaling	Cat#9733S
Anti-CHD4	Abcam	Cat#ab70469
Anti-MTA2	Abcam	Cat#ab8106
Anti-LDB1	Santa Cruz Biotechnology	Cat#sc-365074
Chemicals, peptides, and recombinant proteins
Protein A agarose beads	ThermoFisher	Cat#15918014
Protein G agarose beads	ThermoFisher	Cat#15920010
Power SYBR Green PCR Master Mix	ThermoFisher	Cat#4367660
AMPure XP for PCR Purification	Beckman Coulter	Cat#A63880
DpnII	NEB	Cat#R0543
Csp6I	Thermo Fisher	Cat#ER0211
T4 DNA Ligase	NEB	Cat#M0202L
Critical commercial assays
QIAGEN PCR Purification Kit	QIAGEN	Cat#78840
QIAGEN RNeasy Kit	QIAGEN	Cat#28106
TruSeq ChIP Sample Preparation Kit	Illumina	Cat# IP 202–1012
Phusion High-Fidelity PCR Master Mix	ThermoFisher	Cat#F531S
Cell Line Nucleofector Kit R	Lonza	Cat#VVCA-1001
In-Fusion” HD Cloning Plus	Clontech	Cat#638909
Micro Bio-Spin P-30 Gel Columns, Tris Buffer (RNase-free)	Bio-Rad	Cat#7326250
Qubit dsDNA HS Assay Kit	ThermoFisher	Cat#Q32854
Expand™ Long Template PCR System	Millipore Sigma	Cat#11759060001
High-Capacity cDNA Reverse Transcription Kit	Thermofisher	Cat#4368814
Stemcell EasySep Mouse Hematopoietic Progenitor Cell Isolation Kit	Stem Cell Technologies	Cat#19856
Deposited data
ChIP-seq G1E-ER4	This paper	GSE193740
PRO-seq G1E-ER4	This paper	GSE193741
4C-seq G1E-ER4	This paper	GSE193718
ChIP-seq mouse fetal liver	This paper	GSE223244
CUT&RUN mouse fetal liver	This paper	GSE171385
H3K27me3 ChIP-seq CD71+Ter119+ cells	Pimkin et al., 2014	GSE49664
H3K27me3 ChIP-seq G1E and G1E-ER4	Wu et al., 2014	GSE51338
RNA-seq	Jain et al., 2015	GSE40522, GSE51338, and GSE49847
GATA1 ChIP-seq	Jain et al., 2015	GSE51338, GSE36029, and GSE49847
Experimental models: Cell lines
G1E	Mitchell J. Weiss lab	Weiss et al., 1997
G1E-ER4	Mitchell J. Weiss lab	Weiss et al., 1997
G1E-ER4 Kit cDNA overexpressing (G1E-ER4k)	This paper	N/A
Oligonucleotides
ChIP-qPCR primers	This paper	Methods details
RT-qPCR primers	This paper	Methods details
4C primers	This paper	Methods details
CRISPR guide RNAs	This paper	Methods details
Recombinant DNA
pX330-mCherry	Wu et al., 2013	Addgene #98750
MiGR-Kit-cDNA	Gerd A. Blobel lab	N/A
Software and algorithms
R	(R Core Team, 2014)	http://www.R-project.org/
BedTools	Quinlan, 2014	https://bedtools.readthedocs.io/en/latest/
DESeq2	Love et al., 2014	https://bioconductor.org/packages/release/bioc/html/DESeq2.html
Samtools	Li et al., 2009	http://samtools.sourceforge.net/
FLASH2	Magoc and Salzberg, 2011	http://www.cbcb.umd.edu/software/flash
Seqtk	https://github.com/lh3/seqtk/blob/master/	https://github.com/lh3/seqtkREADME.md
4C mapping	Geeven et al., 2018	https://github.com/deWitLab/4C_mapping
peakC	Geeven et al., 2018	https://github.com/deWitLab/peakC
MACS2	Zhang et al., 2008
Other
PRO-seq data analysis	This paper	https://github.com/zhezhangsh/PROseqR