Abstract
Transcriptional enhancers enable exquisite spatiotemporal control of gene expression in metazoans. Enrichment of mono-methylation of histone H3 lysine 4 (H3K4me1) is a major chromatin signature that distinguishes enhancers from gene promoters. Lysine Specific Demethylase 1 (LSD1, aka KDM1A), an enzyme specific for demethylating H3K4me2/me1, has been shown to “decommission” stem cell enhancers during the differentiation of mouse embryonic stem cells (mESC). However, the roles of LSD1 in undifferentiated mESC remain obscure. Here, we show that LSD1 occupies a large fraction of enhancers (63%) that are primed with binding of transcription factors (TFs) and H3K4me1 in mESC. In contrast, LSD1 is largely absent at latent enhancers, which are not yet primed by TF binding. Unexpectedly, LSD1 levels at enhancers exhibited a clear positive correlation with its substrate, H3K4me2 and enhancer activity. These enhancers gain additional H3K4 methylation upon the loss of LSD1 in mESC. The aberrant increase in H3K4me at enhancers was accompanied with increases in enhancer H3K27 acetylation and expression of enhancer RNAs (eRNAs) and their target genes. In post-mitotic neurons, loss of LSD1 resulted in premature activation of enhancers and genes that are normally induced after neuronal activation. These results demonstrate that LSD1 is a versatile suppressor of primed enhancers, and is involved in homeostasis of enhancer activity.
Introduction
Transcriptional enhancers were discovered as potent gene regulatory elements that act independently of the distance and orientation to the target promoters 1, 2. A broad range of physiological and developmental processes rely on coordinated actions of transcriptional enhancers to achieve cell type-specific and temporally-controlled gene expression 3, 4. Numerous non-coding variants associated with a variety of human traits have been observed at enhancers, implicating their importance in normal physiology and disease pathogenesis 5-7.
Recent studies have begun to reveal how the life cycle of enhancers progresses to induce gene expression changes during development 8. Genome-wide discovery of thousands of potential enhancer elements has been facilitated by profiling 1) binding of pioneer transcription factors (TFs), 2) chromatin accessibility as measured by hypersensitivity to DNase I, and 3) patterns of histone modifications (reviewed in 9). Monomethylation of histone H3 lysine 4 (H3K4me1) can distinguish enhancers from promoters, which in contrast are modified with H3K4me3 10, 11. In response to various environmental and developmental cues, TFs bind to specific DNA elements 12, 13, and subsequent recruitment of methyltransferases KMT2C and KMT2D (aka MLL3 and MLL4, respectively) leads to H3K4me1 at enhancers 14-16.
Once installation of H3K4me1 “primes” enhancers, they can become either “active” or “poised”, depending on the acetylation or tri-methylation of H3 lysine 27 (H3K27ac or H3K27me3) 17, 18, respectively. A recent report showed that active enhancers can be negatively regulated by RACK7-mediated recruitment of KDM5C, an H3K4me3/2 demethylase 19. When genes need to be turned off, e.g. pluripotency genes during differentiation of stem cell, their enhancers undergo “decommissioning” by LSD1-mediated removal of the priming mark, H3K4me1 20. Notably, “latent” enhancers, i.e. DNA elements that lack TF binding or H3K4me1, can gain enhancer-like marks in response to extra-cellular stimuli and promote gene expression in fully-differentiated macrophages 21. These findings highlight the dynamic H3K4 methylation of an enhancer during its life cycle.
Two H3K4 demethylases, LSD1 and KDM5C, have been shown to play important roles in regulation of enhancers 19, 20, 22. While KDM5C reverses H3K4me3/2 leaving H3K4me1 intact 23, 24, LSD1 can demethylate only H3K4me2/1 25. The distinct substrate specificities raise a possibility that these two H3K4 demethylases may cooperate to generate and/or maintain the balance of H3K4me landscape at different classes of enhancers. For example, absence of H3K4 methylation at latent enhancers could potentially be attributed to LSD1-mediated demethylation of H3K4me1. Besides the decommissioning of stem cell genes and enhancers during differentiation, LSD1 has also been shown to repress developmental genes 26, 27 and retrotransposons 28 in ES cells. However, it remains unclear whether LSD1 and KDM5C play any role at other classes of enhancers.
In the present study, we demonstrate that in addition to active enhancers, LSD1 also occupies poised enhancers, some of which are quickly activated (inducible enhancers) upon differentiation of mESC or depolarization of post-mitotic neurons. Interestingly, LSD1 does not bind to latent enhancers, e.g. neuron-specific enhancers that are unprimed in mESC, and LSD1 occupancy shows a clear positive correlation with its substrate, H3K4me2. Loss of LSD1, but not KDM5C, leads to a global upregulation of enhancer RNAs accompanied with increased H3K4 methylation and H3K27ac at active, poised, and inducible enhancers, and their target genes. These results indicate that LSD1 is a pervasive suppressor of primed enhancers, involved in negative-feedback mechanisms to maintain homeostasis of histone-modification landscapes at enhancers.
Results
LSD1 occupies a large fraction of primed enhancers
To study the role of LSD1 in regulation of enhancers and gene expression, we first examined the genome-wide distribution of LSD1 at various regulatory elements. We analyzed the previously published ChIP-Seq datasets of LSD1 20, p300 29, 30, CTCF 29, 31, DNase-Hypersensitivity (DHS) 32 and other histone modifications (see Supplementary Table 1). p300, a histone acetyltransferase and a transcriptional coactivator, has been shown to occupy both promoters and enhancers 10, 33, whereas CTCF binding sites anchor chromatin loops 34 and insulating domains 35-37. By examining the overlaps of binding sites of LSD1 (109,541, q < 0.05), p300 (86,426, q < 0.01), CTCF (58,899, p < 10−12) and DHS sites (299,799, q < 0.05), we found that 1) a majority of p300 binding sites (70.5%, Figure 1a) were co-occupied by LSD1, 2) in contrast, only 14.7% of non-p300 CTCF-binding sites were occupied by LSD1, and 3) most of the LSD1 binding sites (86%) showed an overlap (± 250 bases) with DHS sites. The higher degree of overlap of LSD1 with p300 compared to CTCF-only sites was observed at promoter, genic, and intergenic regions (Supplementary Figure 1). These observations indicate that LSD1 occupies a large fraction of primed enhancers.
Next, we sought to identify regulatory elements that could potentially act as enhancers. Previous studies have utilized a high H3K4me1:me3 ratio, either alone 10, 17 or in conjunction with TF binding 38, 39, DHS or binding by CBP/p300 40, 41, to distinguish enhancers from promoters in a given cell-type. H3K4me2 is observed at both promoters and enhancers and has been shown to be a signature to predict enhancers 38. We therefore included H3K4me2 data to increase the sensitivity and precision of enhancer mapping. Enhancers also differ from promoters in that promoters are associated with stable transcripts 42 while active enhancers are associated with expression of enhancer RNA transcripts (eRNAs) 42-44, which are short-lived due to exosome-mediated degradation 44-46. This degradation of nascent transcripts from enhancers results in a very low, albeit, detectable levels of eRNAs in RNA-Seq (Figure 1b, right panel). Global Run-On, an in vitro assay, followed by high-throughput sequencing (GRO-Seq) 47 enables a sensitive and quantitative evaluation of transcriptionally-engaged RNA polymerase molecules. GRO-Seq, thus, serves as an indirect measure of nascent transcription at promoters and enhancers, irrespective of the subsequent stability of the transcripts 42. Therefore, we employed a high ratio of GRO-Seq:RNA-Seq signals to further refine the prediction of enhancers in mESC. We focused on only intergenic enhancers as we found it difficult to differentiate the eRNAs from gene-coding and promoter-upstream 46 transcripts. In summary, intergenic enhancers were defined as ± 500 base regions around p300/DHS summits with i) H3K4me1 enrichment (rpkm ≥ 1 and ChIP:Input > 1.5), ii) H3K4me3 lower than either H3K4me1 or H3K4me2, iii) a low rate of transcription (RNA-Seq fpkm < 0.5), iv) a GRO-Seq:RNA-Seq ratio > 5, and v) a high average mappability to exclude repetitive regions. This pipeline predicted a total of 22,047 intergenic enhancers in mESC (Supplementary Table 2).
LSD1 has been shown to occupy enhancers in various cell types 20, 48, 49. However, the genome-wide relationship between LSD1 binding and chromatin states at enhancers, such as histone-modification landscapes and eRNA transcription, remains unclear. To address this issue, we first subdivided the 22,047 predicted intergenic enhancers into quartiles (Q1-Q4, Figure 1b) based on the enrichment of H3K4me2 relative to H3K4me1. Similar to a previous observation in K562 cells 42, we noted a positive correlation between eRNA levels, measured by GRO-Seq, Nuclear RNA-Seq or RNA-Seq, and H3K4me2 levels (Figure 1b). Acetylations of H3K9 and H3K27 and eRNA expression have been established as signatures of active enhancers 9, 17, 18. Consistently, we also observed that enhancers with higher transcription levels displayed higher acetylation levels of H3K9 and H3K27 relative to trimethylation (Supplementary Figure 2).
We then determined the extent of LSD1 binding across various enhancer classes and found that a large fraction (63.2%) of the predicted 22,047 enhancers is bound by LSD1 (Figure 1c). Surprisingly, we found that LSD1 occupancy at enhancers increased with increasing levels of H3K4me2 or increasing levels of GRO-Seq signals (Supplementary Figure. 3 and 4a). We then calculated the correlation coefficients between LSD1 levels and levels of various histone modifications at promoter distal regions, i.e. excluding TSS ± 1500 bases. Compared to H3K4me1 (r = 0.5745) and H3K4me3 (r = 0.486), we found that LSD1 levels showed the highest correlations with its primary substrate, H3K4me2 (r = 0.627, Supplementary Figure 4b), and H3K27ac (r = 0.604), a marker for enhancer activity. In contrast, H3K27me3 and H3K9me3 were inversely correlated with LSD1 levels (Supplementary Figure 4b). We also classified 22,047 intergenic enhancers into poised, active, and intermediate enhancers based on their H3K27me3 and H3K27ac levels, according to the previous report 18. We found that LSD1 occupies substantial fractions of each of the three enhancer classes with increased occupancy of active enhancers compared to other classes (Figure 1d). Similar patterns were observed when we used the levels of H3K9me3 and H3K9ac to classify enhancers as described previously 41(Figure 1e). These results indicate that LSD1 binds to multiple enhancer classes with positive correlations to H3K4me2, H3K27/K9 acetylation and eRNA levels.
LSD1 rarely binds to cell-type specific “latent” enhancers
Higher LSD1 occupancy at more active enhancers may contradict with the LSD1’s classical role as a transcriptional repressor of neuron-specific genes in non-neuronal tissues 25, 50. RE1-Silencing Transcription factor (REST) is known to be expressed in non-neuronal cells with the role of repression of neuronal genes in these cell-types through the Corepressor of REST(CoREST) complex 50. We analyzed the previously published REST ChIP-Seq dataset 20 and found that only 1.31% of predicted enhancers (288 out of 22,047) were bound by REST in mESC. However, a majority (91%) of these REST-positive enhancers were bound by LSD1. These data suggest that REST/LSD1-mediated suppression is not the primary mechanism of enhancer regulation in mESC. Enhancers that are inert, free of TFs, and thus, insensitive to DNase I, are referred to as “latent” enhancers in a given cell type 21. The low occupancy of primed enhancers by REST/LSD1 prompted us to test if LSD1 contributes to regulation of latent enhancers. We first looked at the mir290 cluster, which is specifically expressed in mESC and early developmental stages 51. In mESC, several enhancers upstream of its promoter, show high DHS, p300 binding and high GRO-Seq signals accompanied with strong LSD1 binding events (Figure 2a). These mir290 enhancers lack brain-derived DHS, GRO-Seq signal and H3K4 methylation in cortical neurons (CN), demonstrating that these enhancers are primed in mESC but are latent in CN. LSD1 ChIP-Seq data from neural stem cells (NSC) 52, however, showed a lack of LSD1 binding at these latent enhancers. Conversely, two enhancers upstream of Npas4 (marked with asterisks, Figure 2b), a gene predominantly expressed in the brain, showed brain-specific DHS and LSD1 occupancy, GRO-Seq signals and high H3K4me1, specifically in the neuronal cell types (NSC or CN). In mESC, LSD1 is absent at these brain-specific DHS sites upstream of the Npas4 promoter (Figure 2b). These two examples suggest that LSD1 could primarily be recruited to primed enhancers in a given tissue in a TF-binding dependent manner.
To ascertain this specificity of LSD1 recruitment to primed enhancers on a genome-wide scale, we sought to identify genomic elements that are latent in mESC but are primed in other cell types. We identified DHS sites from mESC (398,675, q < 0.01) and four additional mouse tissues, including adult brain (415,400), heart (320,416), liver (207,046), and lung (358,575), using Hotspot (v4.1) 53. Similar to our earlier observation (Supplementary Figure 1), we found that most of mESC LSD1-binding sites (86%) overlapped with mESC hotspots (Figure 2c). Next, we performed an intersection of hotspots from the five tissues. This resulted in tens of thousands of hotspots, which could potentially act as tissue-specific enhancers in a given tissue and latent in others (Figure 2d, e). Motif analysis on promoter-distal hotspots revealed that these tissue-specific hotspots are indeed enriched with binding sites for lineage-specific TFs (Supplementary Figure 5). In agreement with the mir290 and Npas4 loci, mESC LSD1 binding sites showed negligible overlaps with tissue-specific hotspots (0.40-0.69%), whereas 14.31% of mESC-specific hotspots were bound by LSD1 (Figure 2d, e). Based on these data, we concluded that LSD1 is predominantly recruited to primed enhancers and is not actively involved in maintaining inactivity of latent enhancers in mESC.
Loss of LSD1 results in a genome-wide increase in enhancer H3K4 methylation and H3K27 acetylation
The unexpected positive correlation between LSD1 binding and enhancer H3K4me2 levels raises a possibility that LSD1 may not be demethylating H3K4me2 at enhancers. LSD1 has been implicated in demethylation of H3K9 instead of H3K4 when it binds androgen receptors 54, 55, though a recent study reported otherwise 56. Phosphorylation of H3T6 appears to interfere with LSD1-mediated H3K4 demethylation 57. Alternatively, the positive correlation may reflect a negative feedback mechanism, in which LSD1 searches for and binds to genomic regions with high H3K4me2 levels and reverses this modification to regulate optimal enhancer activity.
To test whether LSD1 is involved in maintaining precise levels of H3K4 methylation at enhancers, we investigated the previously generated mESC line that lacks LSD1 due to the insertion of a gene-trap cassette (Lsd1-GT) 28. Western blot analysis of mESC carrying either wild-type (WT) Lsd1 or Lsd1-GT did not show any detectable differences in total H3K4me1, H3K4me2, H3K4me3 or H3K27ac levels (Supplementary Figure 6). We then performed ChIP-Seq to measure H3K4me levels across the genomes of these two mESC lines. Since LSD1 is known to associate with multiple HDAC-containing co-repressor complexes, including the CoREST 50, 58 and NuRD 59 complexes, we also included H3K27ac and HDAC1 in our ChIP-Seq analysis.
Genome-wide localization analysis (ChIP-Seq) profiles, which reflect the spatial distribution of these marks, looked highly similar between the two genotypes at most of the loci. Upon the loss of LSD1, however, H3K4 methylations displayed statistically-significant increases at active, poised, and intermediate LSD1-target enhancers, which were accompanied with conspicuous increases in H3K27ac (Figure 3a, Supplementary Figure 7a). Similar changes were also observed at enhancers that showed a significant increase in eRNA expression (see next section) in the Lsd1-GT mESC (Supplementary Figure 7). Interestingly, HDAC1 levels did not change significantly at poised enhancers, while active or poised enhancers showed a small but significant increase in HDAC1 binding (Figure 3a), which could be attributed to either experimental variations or unknown mechanisms to compensate for the loss of LSD1. The inability of HDAC1 to remove H3K27ac in Lsd1-GT cells is consistent with the previous observations that HDAC activity is negatively influenced by the presence of H3K4me 27, 60. Representative genes Pou5f1 (Figure 3b) and Cbln4 (Figure 3c), which are normally active or poised in undifferentiated mESC, respectively, showed relatively higher H3K4me and H3K27ac at both promoters and enhancers in Lsd1-GT mESC. We also observed a concomitant increase in both promoter- and enhancer-associated GRO-Seq signals at these loci in Lsd1-GT mESC (Figure 3b, c). The increases in H3K27ac and nascent transcription at these poised and intermediate enhancers upon the loss of LSD1 suggest a shift in their identity towards active enhancers. These results indicate that LSD1 functions as an H3K4 demethylase at enhancers, and is required for maintenance of optimal H3K4me and H3K27ac levels in mES cells.
Loss of LSD1 but not KDM5C results in aberrant activation of transcriptional enhancers
In the above ChIP-Seq study, we found a genome-wide elevation of all three H3K4me statuses, including H3K4me3, at LSD1-target enhancers in Lsd1-GT mESC (Figure 3a, Supplementary Figure 7a). This increase of H3K4me3 at enhancers cannot be explained directly by the loss of LSD1, as LSD1 is incapable of demethylating H3K4me3 25; therefore, one or more H3K4me3 demethylases might be involved in maintaining low levels of H3K4me3 at enhancers. LSD1 and KDM5C, an H3K4me3/me2 demethylase 23, have been previously shown to be in the same complex 19 and that KDM5C suppresses over-activation of active enhancers in breast cancer cells 19. To elucidate the interrelationship between LSD1 and KDM5C in suppression of enhancer activity, we first performed KDM5C ChIP-Seq in mESC and identified 113,166 KDM5C binding sites (MACS2, q < 0.05). Most of the 22,047 predicted intergenic enhancers (78.3%) were bound by either LSD1 or KDM5C and 52.1% of total were bound by both (Figure 4a).
We generated Kdm5c-knockout (KO) mESC by transfecting a Cre-expression plasmid into the mESC harboring the floxed exons 11 and 12, which encode the catalytic JmjC domain 61, 62, and confirmed the loss of KDM5C (Supplementary Figure 8). We then asked if the loss of either LSD1 or KDM5C leads to aberrant enhancer activity by quantifying changes in GRO-Seq signals at enhancers. To identify misregulated enhancers, we calculated the number of GRO-Seq reads mapping within ± 500 bases of the center of the predicted enhancers and normalized them against 199,209 p300/DHS sites across the whole genome using DESeq 63. Upon the loss of LSD1, a large fraction (24.8%, 5,471) of total intergenic enhancers showed a significant elevation in associated GRO-Seq transcripts, while a small number 674 (3.06%) displayed a reduced activity with a stringent cutoff of q < 0.05 (Figure 4b). Next, we tested if this elevation of GRO-Seq signals is specific to poised, intermediate or active enhancers. We found that all three enhancer classes showed a significant increase in associated nascent transcripts (p < 2.2e-16, Wilcoxon signed-rank test, Figure 4c), indicating that LSD1 is required for genome-wide suppression of aberrant enhancer activities.
Using the same DESeq cutoff, however, we were not able to identify any misregulated enhancers in Kdm5c-KO mESC. After relaxing the cutoff to p < 0.05, we could identify only 63 upregulated and 102 downregulated enhancers upon the loss of KDM5C (Figure 4d). To confirm that these observations were not dependent on differences in sequencing depths or inter-replicate variability, equal number of reads were randomly selected from each GRO-Seq sample and pairwise DESeq comparisons between individual replicates of either genotypes were repeated. Thus, in contrast to the crucial role of LSD1, KDM5C is largely dispensable for enhancer suppression in mESC.
Since GRO-Seq is an in vitro transcription assay, we sought to validate this global upregulation of enhancers upon LSD1 depletion in mESC under physiological conditions by sequencing total cellular RNAs (RNA-Seq) and nuclear RNAs (Nuclear RNA-Seq). Either RNA-Seq or Nuclear RNA-Seq could not provide sufficiently high eRNA signals to call differentially-expressed enhancers likely due to the aforementioned exosome-mediated degradation of eRNAs. However, when we evaluated eRNA levels at all the intergenic enhancers as a group, RNA-Seq and Nuclear RNA-Seq corroborated our GRO-Seq results (Supplementary Figure 9). These data demonstrate that LSD1, but not KDM5C, is required for suppression of aberrant enhancer activities in mESC.
Aberrant changes in enhancer activity are associated with misregulation of physically-interacting genes
The standard approach to gauge the influence of enhancer misregulation on gene expression has been to quantify changes in expression of genes that are located in proximity to the enhancers of interest. However, recent advances in genome-wide profiling of chromatin interactions 64-66 have paved the way for a more precise determination of enhancer-promoter interactions. To identify genes that physically interact with our set of predicted enhancers, we utilized the recently-published “HiCap” data set, which is a high-resolution map of promoter-anchored chromatin interactions in mESC 67. For instance, our enhancer prediction identified a ~ 6 kb-wide enhancer cluster downstream of Dusp5 and the analysis of HiCap data revealed that one of the three individual enhancers within the cluster appears to interact with the Dusp5 promoter (Figure 5a). This enhancer cluster was significantly upregulated in Lsd1-GT mESC and showed a pronounced increase in H3K4me2 and H3K27ac levels, and Dusp5 transcription (Figure 5a). A concomitant misregulation of enhancers and the interacting gene was also observed for the aforementioned mir290 cluster (Supplementary Figure 10a).
To evaluate the genome-wide impact of enhancer misregulation on gene expression, we first categorized enhancers based on the statistical significance of their differential expression in Lsd1-GT mESC: significantly misregulated (q < 0.05) and moderately misregulated (0.05 ≤ q < 0.25) enhancers from our GRO-Seq analysis. We then retrieved the physically-interacting promoters from the HiCap data, and plotted the changes in mRNA levels (RNA-Seq, Figure 5b) or rates of nascent transcription (GRO-Seq, Supplementary Figure 10b). We observed a general trend that genes associated with upregulated enhancers showed an increased expression and vice versa, and the genes that are anchored to unaffected enhancers did not exhibit any significant changes in expression in Lsd1-GT mESC (Figure 5b, Supplementary Figure 10b, c). Importantly, for each category, the magnitude and the statistical significance of median change in gene expression correlated positively with those of changes in enhancer activity (Supplementary Figure 10c). Additionally, when interacting genes were called on the basis of genomic proximity to the enhancers, we observed a similar trend (Supplementary Figure 11). These results indicate that LSD1’s role at enhancers is important for a precise transcription of their cognate genes.
To further corroborate if LSD1 and its catalytic activity are required for suppression of enhancer activity and associated genes, we utilized luciferase reporter assays in Lsd1-GT mESC. We selected 11 enhancers with two latent enhancers and at least one enhancer from each of poised, intermediate, and active enhancers that showed significant upregulation in Lsd1-GT mESC compared to WT-mESC and also showed an upregulation of the associated gene. 1.0 ~ 1.2 kb of the enhancer-containing regions were cloned downstream of the HSV-Thymidine Kinase promoter-driven firefly luciferase gene. Lsd1-GT mESC were transfected with a control plasmid or plasmids expressing either human LSD1 or the catalytically inactive LSD1-K661A mutant 68 along with the reporter plasmids. We found that LSD1’s catalytic activity is indeed required for suppression of all the active enhancers tested (p < 0.1, Student’s t-tests, Supplementary Figure 12), consistent with high levels of LSD1 at these enhancers in mESC. In contrast, one of the Nanog enhancers, which had not displayed a change upon the loss of LSD1 in mESC, was unaffected by LSD1 expression. We observed lower enhancer activities of the latent, poised, and intermediate enhancers compared to the active enhancers, indicating that our enhancer classification could accurately predict enhancer activity. However, we found it difficult to interpret the effect of LSD1 or its catalytic activity at these weak enhancers as they failed to enhance the activity of the promoter.
Both mESC-specific and differentiation genes are upregulated upon the loss of LSD1 in undifferentiated mES cells
A previous study has implicated LSD1 in decommissioning of enhancers of pluripotency genes during differentiation of mES cells 20. However, the roles of LSD1 in undifferentiated mES cells remain elusive. The global upregulation of enhancers (Figure 4b), prompted us to investigate if the loss of LSD1 in undifferentiated mES cells affected the expression of pluripotency and/or differentiation genes. To this end, we first analyzed a published RNA-Seq datasets for mESC and epiblast stem cells (EpiSC), which were derived from mESC by treatment with Activin A and FGF2 for 4 days 69. We first selected the genes that showed a significant (q < 0.01, DESeq) and at least 5-fold change in expression during differentiation. This analysis yielded 710 induced and 745 downregulated (decommissioned) genes during the differentiation of mESC to EpiSC (Figure 6a). mESC and EpiSC represent two consecutive stages of embryonic development, namely pre-implantation and post-implantation, respectively. Therefore, these sets of upregulated and downregulated genes may represent the earliest transcriptional response of mES cells to the differentiation cue.
Compared to WT mESC, Lsd1-GT mESC displayed roughly equal number of genes being significant-upregulated or -downregulated (55% vs. 45%, 1493 upregulated and 1203 downregulated, respectively, q < 0.05). Quantitation of gene expression changes of the induced and the decommissioned genes, using our RNA-Seq and GRO-Seq datasets revealed that many of these genes are upregulated in undifferentiated Lsd1-GT mESC and both of these gene sets showed statistically-significant upregulation as group (p < 2.2e-16, Wilcoxon signed-rank test, Figure 6b, c). For example, Hmga2 is a gene which is induced upon differentiation of mES cells and is required for the exit from naïve pluripotency 70. As shown in Figure 6d, LSD1 loss led to a marked increase in Hmga2 expression which was also associated with increased H3K4 methylations, H3K27ac, and GRO-Seq signals at nearby enhancers. We have shown earlier that some pluripotency genes including Pou5f1 (Figure 3b) and mir290 (Supplementary Figure 10a) and their nearby enhancers were upregulated in Lsd1-GT mESC. Upregulation of both pluripotency and differentiation genes upon the loss of LSD1 in mES cells suggests that LSD1 does not instruct the fate of mES cells to a particular direction.
LSD1 is required for suppression of inducible enhancers in terminally-differentiated neurons
We next sought to test if LSD1-mediated regulation of enhancers plays a role in gene expression program of terminally-differentiated cells using cortical neuron (CN) culture as a model. Using lentiviral delivery of two independent short hairpin RNAs (shRNAs) at 7 days in vitro (DIV), we knocked down (KD) LSD1 in primary cultures of mouse CN (Figure 7a). We employed BrU-Seq, a recently-developed nascent-RNA sequencing technique 71, that allows an accurate evaluation of any changes in active transcription of both mRNAs and eRNAs. At DIV 11, i.e. after 4 days of control or Lsd1 shRNA delivery, neuronal cultures were treated with 5-bromouridine for 32 min, followed by enrichment of BrU-containing nascent transcripts using anti-BrdU beads and high-throughput sequencing. DESeq analysis indicated a significant misregulation of 1,500 genes (q < 0.05, 778 downregulated and 722 upregulated). Interestingly, many well-characterized activity-regulated genes (ARGs) 43, including Arc, Fos, Fosb, Npas4, Egr1-4, and Nr4a1-3, were among the most significantly-upregulated genes upon LSD1-KD in unstimulated neurons (Figure 7b). ARGs are expressed at low levels in resting neurons and are rapidly induced by depolarization of neurons via sensory inputs, thereby representing a stimulus-responsive gene regulatory program. Since products of ARGs play important roles in synaptic plasticity underlying cognitive development, learning and adaptive processes 72-74, we narrowed our focus on these inducible genes.
To evaluate this ARG upregulation on a genome-wide scale, we analyzed our previously published RNA-Seq data set 75 and identified 140 ARGs that were induced by KCl-mediated depolarization of CN in culture. Both of the two independent Lsd1 shRNAs led to a spurious induction of many ARGs in the resting neurons (Figure 7b, Supplementary Figure 13), indicating that LSD1 suppresses premature induction of ARGs in CN. We next examined the Npas4 locus to check if enhancer misregulation upon the loss of LSD1 could be involved in the premature induction of ARGs. We identified three putative enhancers upstream of the Npas4 promoter based on DHS and H3K4me1 enrichment (Figure 7c). These three enhancers appear to respond to membrane depolarization, as they show activity-dependent increases in NPAS4 binding 43 and H3K27ac levels 76. These activity-regulated Npas4 enhancers are bound by LSD1 in NSC 52. An increase in nascent transcription across these enhancers, concomitant with an increased Npas4 expression, indicates that these enhancers are deregulated upon the loss of LSD1 (Figure 7c). A previous study had reported more than ten thousand putative activity-regulated enhancers based on increased CBP binding in response to membrane depolarization 43. Subsequent work categorized these candidate enhancers into four groups on the basis of activity-dependent changes in H3K27ac 76 (Figure 7d). The study found that only the enhancers that displayed activity dependent changes in H3K27ac were involved in promoting ARG transcription 76. Next, we investigated if the premature upregulation of ARGs in Lsd1-KD neurons was accompanied with misregulation of any of these activity-regulated enhancer groups. Analysis of BrU-Seq data revealed that loss of LSD1 did not have a significant impact on enhancers that do not display any activity-dependent changes in H3K27ac (Wilcoxon signed-ranked test, Figure 7d). However, Lsd1-KD led to a significant upregulation of eRNA levels at the enhancers that gain or lose H3K27ac upon KCl treatment (Figure 7d, Supplementary Figure 14a). Interestingly, the group of enhancers with no H3K27ac either before or after depolarization, and are presumably poised neuronal enhancers, also showed an upregulation upon the loss of LSD1 (Figure 7d, Supplementary Figure 14a). Similar to our earlier observations with RNA-Seq and Nuclear RNA-Seq in mESC, the eRNA signals with BrU-Seq were considerably lower than those with mESC GRO-Seq to obtain sufficiently-high statistical power for comparison; therefore, we aggregated eRNA signals from the two control and the two Lsd1-KD experimental groups for this analysis (Figure 7d, Supplementary Figure 14a). Similar trends were observed in analysis without grouping the samples (Supplementary Figure 14b). These data indicate that LSD1 is required for genome-wide suppression of premature enhancer activation in resting neurons.
Activation of ARGs upon Lsd1-KD could also be a result of extraneous activation of signaling pathways upstream of ARG induction. Extracellular signal-regulated kinases 1 and 2 (ERK1/2) are rapidly phosphorylated in response to a variety of extracellular stimuli, including membrane depolarization, and play critical roles to mediate the transcriptional response 72, 77. A lack of noticeable differences in phosphorylation levels of ERK1/2 upon Lsd1-KD (Supplementary Figure 14c), further support a direct role of LSD1 in suppression of activity-regulated enhancers and genes.
Discussion
Early embryonic lethality of homozygous Lsd1-KO mice indicates an essential role of LSD1 in development 78. However, the roles of LSD1 in early embryogenesis have not been fully elucidated. A previous study has shown that the silencing of pluripotency genes during differentiation of mES cells is mediated by the decommissioning of pluripotency enhancers by LSD1 20. This study employed Tranylcypromine (TCP), a pharmacological agent to block LSD1’s enzymatic activity 20. However, TCP also inhibits the H3K4 demethylase activity of LSD2 (aka KDM1B) 79, the paralog of LSD1, which is involved in regulation of transcriptional elongation 80. Thus, it remains unclear whether the observed impact of TCP treatment on enhancer dysregulation in mES cells was mediated by inhibition of either LSD1 or LSD2 or both. By employing genetic ablation of Lsd1 in mES cells, we demonstrate that LSD1 suppresses the activity of a large fraction of primed enhancers, including the pluripotency enhancers and poised enhancers of differentiation genes. Notably, several of these key pluripotency enhancers and genes are already upregulated in undifferentiated Lsd1-deficient mESC (Figure. 3 and 6). Similar to our observations, another group had found that Lsd1-KD by siRNA led to an upregulation of several stem cell genes in undifferentiated mES cells 81. Loss of LSD1 in stem cells has been implicated in multiple differentiation defects, including de-differentiation of the pluripotent mESC state towards the totipotent 2-cell state 82, or the premature differentiation of human ES cells to endodermal and mesodermal lineages 26. These observations could be reconciled by our findings that LSD1 is required for suppressing both pluripotency genes and differentiation genes in mES cells, possibly through maintenance of proper enhancer activity.
We provide several lines of evidence that LSD1 plays an essential role in genome-wide homeostasis of primed enhancers. We show that recruitment of LSD1 correlates positively with levels of enhancer H3K4me2, H3K27ac and eRNA transcription (Figure 1) and this recruitment is specific to primed enhancers (Figure 2). Loss of LSD1 led to an upregulation of a large number of enhancers, as demonstrated by increased H3K4 methylation, H3K27ac (Figure 3), and eRNA transcription (Figure 4), concomitant with an upregulation of the associated genes (Figure 5). These results support the following model of LSD1-mediated homeostasis of the histone modification landscape during the life cycle of an enhancer. Binding of TF and subsequent recruitment of MLL3/4 14-16 prime the enhancers with H3K4me1/me2, which attract LSD1 irrespective of whether the enhancers are destined to be either “active” or “poised” (Supplementary Figure 15). LSD1 then counteracts with MLL3/4 to maintain an optimal H3K4me levels. Enhancers with a relatively low H3K4me2 may represent early stages of priming by TFs and MLLs. It is possible that enhancers with low levels of H3K4me2 recruit little LSD1, which is not detectable by ChIP-Seq (Q1, Figure 1b). When gene expression needs to be increased, recruitment of additional factors and/or MLL3/4 may convert these less active enhancers to more active enhancers with higher H3K4 methylation and H3K27ac, which would then require higher levels of LSD1. LSD1’s recruitment might also serve as a surveillance mechanism to suppress ectopic installation of H3K4 methylation and spurious activation of enhancers.
During differentiation of mES cells, the pluripotency enhancers may be unprimed by the loss of ES-specific TFs followed by the loss of MLL3/4 and H3K4 methylation. Our model is not mutually exclusive to LSD1-mediated decommissioning of enhancers, as LSD1 could remove remnant H3K4me2/me1 to completely disengage the enhancer from active regulation. To further elucidate the mechanisms of decommissioning of pluripotency enhancers, it will be important to determine how early differentiation cues shift the balance of MLL3/4-mediated H3K4 methylation and LSD1-mediated demethylation.
Enrichment of H3K4me1 and depletion of H3K4me3 was the first combination of chromatin signatures to predict a large number of transcriptional enhancers in a mammalian genome 10, 11. More recent studies have shown H3K4me3 to be present at a subset of active enhancers 83 with a positive correlation between the H3K4me3/me1 ratio and enhancer transcription levels 42. We found that KDM5C and LSD1 can co-occupy enhancers in mES cells (Figure 4a). However, only the loss of LSD1, but not KDM5C, displayed significant changes in enhancer activity and gene expression, highlighting an essential and non-redundant role of LSD1 in mES cells. KDM5C has been implicated in both promotion of enhancer activity by generating H3K4me1 in mES cells 22, and suppression of over-activation of enhancers in breast cancer cells 19. Consistent with the former study, our analysis found a small reduction in eRNA levels in Kdm5c-KO mESC (Supplementary Figure 9). In addition to Kdm5c, other KDM5 family members, Kdm5a and Kdm5b, are also expressed in mES cells at similar levels and could possibly compensate for the loss of KDM5C. These observations suggest differential requirement of KDM5C in context of either different enhancers or different cell types.
Repeated “write-and-read” of histone modifications can form a feed-forward loop to allow TF-independent maintenance and propagation of chromatin status. Such models have been well established for the propagation of H3K27me3 84 and H3K9me3 85. Maintenance and propagation of H3K4me epigenetic memory across generations by LSD1 have been observed in worms 86 and mice 87. PHF21A (aka BHC80), another member of the LSD1 complex 50, 58, was the first known reader protein to recognize unmethylated H3K4 88. This unique combination of an H3K4 demethylase and a reader of unmethylated H3K4, makes the LSD1-PHF21A complex an ideal candidate to exert a self-perpetuating “erase-and-read” mechanism. However, a positive correlation between LSD1 and H3K4me2 and LSD1’s absence at latent enhancers suggest that the role of LSD1 in maintaining the epigenetic memory could be limited to other genomic elements and warrants further investigation.
LSD1’s role in suppression of primed enhancers does not appear to be restricted to mES cells. Similar to our observations in mESC, our BrU-Seq analyses in post-mitotic neurons revealed that LSD1 suppresses premature activation of neuronal activity-regulated genes and enhancers (Figure 7). LSD1 has a neuron-specific isoform (neuroLSD1 or LSD1n) with four extra amino acids in the catalytic domain 89 and an altered substrate specificity which remains ambiguous 49, 90. Since our RNAi approach depleted both neuroLSD1 and canonical LSD1 in CN, it remains unclear if one or both of LSD1 isoforms mediate the suppression of activity-regulated enhancers and ARGs. Given that the genetic ablation of neuroLSD1 led to a downregulation of ARG expression 49, 91, it is more likely that the canonical LSD1, and not neuroLSD1, is involved in the suppression of activity-regulated enhancers. Loss-of-function >LSD1/KDM1A >mutations have been genetically associated with several neurodevelopmental conditions 92-94. These disorders could possibly be attributed to uncontrolled activation of activity-regulated genes and enhancers upon the loss of LSD1 and/or neuroLSD1. LSD1-mediated homeostasis of transcriptional enhancers, therefore, underlies various physiological processes including embryonic development and human cognitive function.
Methods
See supplementary methods for details.
Cell culture
Lsd1-WT and Lsd1-GT mESC have been described previously 28. Kdm5c-KO mESC were derived from the previously described mESC that carry the floxed Kdm5c allele 75 by Cre-mediated deletion of exons 11 and 12, which encode the enzymatic JmjC domain. mESC were grown on gelatin coated plates.
Western blot analysis
mESC or CN were lysed in Laemmli sample buffer, sonicated, and subjected to SDS-PAGE. Western blot analyses were carried out using standard protocols using anti-H3K4me1 (ab8895, Abcam), anti-H3K4me2 (ab7766, Abcam), anti-H3K4me3 (ab8580, Abcam), anti-H3K27ac (39135, Active Motif), anti-KDM5C 75, anti-LSD1 (ab17721, Abcam) 20 and anti-phospho-ERK1/2 (4370, Cell Signaling Technology).
Lsd1 knockdown in mouse cortical neurons
Primary cultures of cortical neurons were carried out as described previously 75. Lsd1-KD in CN was achieved by lentiviral delivery of either scramble shRNA (SHC202, Sigma) or Lsd1-shRNAs (A: TRCN0000071375 and B: TRCN0000071376, Sigma) 95 on 7 days in vitro (DIV) and 5-Bromouridine incorporation was performed on DIV 11.
ChIP-Seq
Antibodies used for chromatin immunopreciptation (ChIP) were anti-H3K4me1 (ab8895, Abcam and 07-436, EMD Millipore), anti-H3K4me2 (05-790, EMD Millipore), anti-H3K4me3 (04-745, EMD Millipore), anti-H3K27ac (39135, Active Motif), anti-HDAC1 (A300-713A, Bethyl Laboratories and sc-6298, Santa Cruz Biotechnology) and anti-KDM5C75. KDM5C ChIP-Seq experiments were performed as described previously 75. Other ChIP experiments were performed as described previously 96 with minor modifications.
RNA-Seq and Nuclear RNA-Seq
RNA-Seq libraries have been described in detail previously 97. For sequencing of nuclear RNA, nuclei were isolated as described previously 98 with minor modifications. Libraries from rRNA-depleted RNA were prepared using Direct Ligation of Adapters to First-strand cDNA (DLAF) 97.
Global Run-On
GRO was modified from the method described previously 47, 98. In addition to the presence of 0.2% IGEPAL CA-630, GRO on Kdm5c mESC, Lsd1 mESC and CN were done in presence of 0.5%, 0.25% and 0.2% of >N-Lauroylsarcosine, respectively for 8 min at 30°C.
BrU-Seq
Cortical neurons (DIV 11), after shRNA treatment for four days, were incubated with 2 mM 5-Bromouridine (850187, Sigma) for 32 min at 37°C. To reduce the number of steps for library preparation, we developed Direct Ligation of Adaptor to the 3’ end of RNA (DLAR), a method suitable for preparation of libraries for BrU-Seq.
All sequencing experiments were conducted in biological duplicates concurrently with different genotypes to minimize technical variations.
Sequencing and Alignment
Multiplexed libraries were subjected to single-end sequencing on Illumina HiSeq 2000/2500 instruments using standard oligonucleotides designed for multiplexed paired-end sequencing, except that BrU-Seq indices were sequenced with DLAR_Index_Read:5’-CATAGGAAGAGCACACGTCTGAACTCCAGTCAC-3’.
ChIP-Seq reads were mapped to the mm9 genome using Bowtie1 (v1.1.2)99 allowing for up to two mismatches. PCR duplicates from ChIP-Seq reads were removed using samtools rmdup utility (v1.3) 100 and coverage along the genome was calculated using BEDTools (v2.25.0)101 after extending the ChIP-Seq reads to a total length of 180 bases. RNA-Seq libraries were mapped to the mm9 genome and transcriptome using TopHat2 (v2.1.0)102 with Bowtie2 (v2.2.6) 103. For GRO-Seq and BrU-Seq, full length reads were first aligned using Bowtie1 or Tophat2, respectively. Adaptor sequences were trimmed from the unmapped reads using BBDuk utility 104 and reads were remapped and merged to the reads from the initial alignment. Only uniquely mapping reads were retained for further analysis and libraries were normalized to total number of non-mitochondrial and non-ribosomal reads.
Analysis
MACS2 (v 2.1.0)105 was used to call DHS or ChIP-Seq peaks. For selection of candidate p300/DHS sites for enhancer prediction, we first scanned the genome for the strongest (with highest MACS2 signal) p300 or DHS site in a 1,250 base sliding window. When both p300 and DHS sites were present in the same window, p300 binding site was given higher precedence over any DHS sites. Intergenic p300/DHS sites were defined as sites that were outside of 1.25 kbp upstream to 3 kbp downstream of the genes. LSD1 has been shown to be involved in silencing of repetitive elements including endogenous retroviral elements (ERVs) 28. Therefore, to focus on prototypical enhancers in this study, we excluded p300/DHS sites with a low mappability (M1 < 0.75 and M2 < 0.75), where M1 and M2 indicate the fraction of uniquely mapping bases 106 within ± 500 and ± 100 bases, respectively, of the p300/DHS site. p300/DHS sites within the ENCODE blacklisted regions 5 were also excluded.
FeatureCounts 107 was used for calculating the number of reads overlapping various genomic features. Intersection analyses were done using BEDTools. DESeq (v1.22.1) 63 was used for normalization and differential gene expression analysis. 10-20 genes with exceptionally high expression and miRNAs and were excluded from further analysis. ChIP-Seq enrichment for H3K4me2, H3K4me3 and H3K27ac were normalized using MAnorm 108 with the fraction of reads aligning within the common peaks of WT and Lsd1-GT mESC samples from each replicate. ChIP-Seq coverage profiles from only one replicate, which did not require MAnorm normalization, were used in the browser snapshots in the figures. Prioritization of enhancer assignment is detailed in the supplemental information. Activity-regulated genes were identified as genes showing significant upregulation (p < 0.05, DESeq) in each of the two independent replicates of previously published RNA-Seq datasets from untreated and KCl-treated CN 75. Wilcoxon signed-rank tests were performed after log transformation of changes in expression or ChIP enrichment. The Perl scripts used for analyses are available upon request.
Accession Numbers
Raw and processed sequence data files are available on the Gene Expression Omnibus (GEO) under accession GSE93952. The data can be accessed by the reviewers at: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?token=ulchsiocxvojtaj&acc=GSE93952
Author contributions
SA and BR conceived the project and designed the experiments. SI and SA performed the RNA-Seq, Nuclear RNA-Seq, and GRO-Seq experiments. SA, EB, and TSM performed the ChIP-Seq experiments. SI, PMG, and SA designed and performed the BrU-Seq experiment. SA and RSP performed the luciferase reporter assays. YMN, PMG, and SA performed the western blot analyses. SA established the protocols and performed the computational analysis. SA and SI designed the analysis and wrote the manuscript. All authors approved and edited the manuscript.
Competing financial interests
The authors declare no competing financial interests.
Acknowledgements
We thank Dr. Fulai Jin and Dr. Gary Hon at LICR, San Diego, for helpful discussions and guidance with bioinformatics. We are also grateful to Dr. Nima Mosammaparast at Washington University School of Medicine in St. Louis for the kind gift of WT and catalytically inactive LSD1 >expression plasmids. The work was supported by grants from the LICR (to BR), the California Institute for Regenerative Medicine (RN2-00905, to BR), the University of Michigan Medical School (to SI), Cooley's Anemia Foundation Fellowship (to SI), NIH (NS089896, to SI), the Farrehi research fund (to SI), NSF Graduate Research Fellowship Program (DGE #1256260, to PMG), University of Michigan Career Training in Reproductive Biology (T32HD079342, to RSP), and the Eunice Kennedy Shriver National Institute of Child Health and Human Development DIR (HD008933, to TSM).
References
- 1.↵
- 2.↵
- 3.↵
- 4.↵
- 5.↵
- 6.
- 7.↵
- 8.↵
- 9.↵
- 10.↵
- 11.↵
- 12.↵
- 13.↵
- 14.↵
- 15.
- 16.↵
- 17.↵
- 18.↵
- 19.↵
- 20.↵
- 21.↵
- 22.↵
- 23.↵
- 24.↵
- 25.↵
- 26.↵
- 27.↵
- 28.↵
- 29.↵
- 30.↵
- 31.↵
- 32.↵
- 33.↵
- 34.↵
- 35.↵
- 36.
- 37.↵
- 38.↵
- 39.↵
- 40.↵
- 41.↵
- 42.↵
- 43.↵
- 44.↵
- 45.
- 46.↵
- 47.↵
- 48.↵
- 49.↵
- 50.↵
- 51.↵
- 52.↵
- 53.↵
- 54.↵
- 55.↵
- 56.↵
- 57.↵
- 58.↵
- 59.↵
- 60.↵
- 61.↵
- 62.↵
- 63.↵
- 64.↵
- 65.
- 66.↵
- 67.↵
- 68.↵
- 69.↵
- 70.↵
- 71.↵
- 72.↵
- 73.
- 74.↵
- 75.↵
- 76.↵
- 77.↵
- 78.↵
- 79.↵
- 80.↵
- 81.↵
- 82.↵
- 83.↵
- 84.↵
- 85.↵
- 86.↵
- 87.↵
- 88.↵
- 89.↵
- 90.↵
- 91.↵
- 92.↵
- 93.
- 94.↵
- 95.↵
- 96.↵
- 97.↵
- 98.↵
- 99.↵
- 100.↵
- 101.↵
- 102.↵
- 103.↵
- 104.↵
- 105.↵
- 106.↵
- 107.↵
- 108.↵