Structurally differentiated cis-elements that interact with PU.1 are functionally distinguishable in acute promyelocytic leukemia

Chromatin immunoprecipitation (ChIP) with PU.1-specific antibody followed by deep sequencing was performed for the APL-derived NB4 cells. As shown in Table 1, a total of 15.6 million 35-bp sequence reads were generated, of which 11.8 million (76%) were aligned uniquely and non-redundantly to the human genome (HG18). Based on the false discovery rate (FDR) of 0.1%, a total of 26,907 significantly enriched ChIP regions with a median length of 429 bp were identified (Table 1, Additional file 1: Figure S1 and Additional file 2: Table S1) through MACS [16]. Figure 1A illustrates a representative chromatin region (6p21.33), showing enriched peaks of PU.1 binding. As shown in Figure 1A and Additional file 3: Figure S2, sharp enrichment peaks were found in proximal or/and distal regions of promoters for previously reported PU.1 target genes (e.g., BTK[17], CSF1R[9], CSF2RB[18], ITGAM[12], NCF2[19] and NCF4[20]), and novel target genes of PU.1, such as NFKBIL1, LST1 and AIF1. As a validation, ChIP-quantitative(q)-PCR was then conducted on the binding regions of 8 previously known targets, 9 randomly selected de novo targets and 4 negative controls (Figure 1B), showing results consistent with those of ChIP-seq in this setting. Since PU.1 primarily is considered as a transactivator in myeloid differentiation [21], we then selected eight enriched regions on a random basis and applied them to luciferase assays in 293T cells (Figure 1C). Clearly, PU.1 transactivated these regions, adding further evidence that the ChIP regions identified in this setting represent bona fide functional binding sites of PU.1 in APL cells.

In an attempt to identify potential features associated with PU.1 binding sites, we first compared binding locations to annotated genes based on the UCSC Genome Browser RefSeq Database [22], and found that 14.1% (3,804/26,907) of the binding sites mapped to the promoter regions, 41.9% (11,271/26,907) to the intragenic regions (gene body) and 44.0% (11,832/26,907) to the intergenic regions (including 16.5% upstream enhancers, 9.9% downstream enhancers and 17.5% distal intergenic regions) (Figure 2A), as classified by the recommended criteria (see Materials and methods). These results indicate that the binding spectrum of PU.1 is highly complex and versatile in APL cells. Next, we conducted sequence evolutionary conservation analysis across 27 vertebrate genomes on each of the mapped sections. As shown in Figure 2B, the summit of PU.1 enriched binding regions revealed higher evolutionary conservation than the flanking regions, implicating the biological relevance of PU.1 binding. Moreover, the highest conservation scores were obtained with the promoter-proximal binding sites, supporting the notion that cis-elements in promoter regions are evolutionally conserved [23]. In contrast, the PU.1 promoter-distal binding sites (including those in the intragenic and intergenic regions) revealed relatively low conservation scores, consistent with the idea that cis-elements in enhancers or other distal regulatory elements are relatively dynamic among species [24].

Next, we performed a correlation analysis between the PU.1 binding sites versus total gene number and nucleotide number on each chromosome, respectively. As a control, we conducted the same analysis with the binding sites of a classic promoter-binding factor RNA polymerase II (RNAPII), a typical enhancer-binding protein estrogen receptor (ER) [25] and some other factors including CTCF [26], STAT1 [27], FOXA1 [28], GATA1 [29] and GATA2 [30] (Figure 2C and Additional file 4: Table S2), whose genome-wide binding sites have been documented previously. Interestingly, the PU.1 binding sites were correlated with both the chromosomal gene number (r ² =0.70) and nucleotide number (r ² =0.72) (Figure 2C), which appeared to be distinguishable from most of the other tested factors except for STAT1, a known promoter-enhancer dual binding TF [31, 32]. The factors like RNAPII, CTCF, GATA1 and GATA2, were obviously correlated with the number of genes, whereas those like ER and FOXA1, known as two specific enhancer-binding proteins, were correlated with the number of nucleotides. The above observations suggest that PU.1 may act as a versatile factor able to interact with cis-elements not only in promoter regions but also in enhancer regions.

In addition, we investigated the PU.1 binding locations and numbers on their corresponding RefSeq genes (9,556), revealing that 33.3% (3,184/9,556) of the RefSeq genes harbored the binding sites on their promoter-proximal regions whereas 66.7% (6,372/9,556) contained the binding sites on the promoter-distal regions. These observations, together with data shown in Figure 2A, suggest that promoter-distal binding of PU.1 may play at least as equally important roles as the promoter-proximal binding in transcriptional regulation. Interestingly, more than half of the genes (1,598/3,184) with promoter-proximal binding of PU.1 appeared to contain additional binding sites in their promoter-distal regions, suggesting that PU.1 regulatory mechanisms can be far more complex than previously recognized, by involving multiple trans-cis interaction sites. This would allow for precise control of gene expression that is essential for myeloid differentiation. Indeed, auto-regulation of the PU.1-encoding gene SPI1 appears to require PU.1 binding at both the promoter [21] and enhancer (i.e., 17 kb upstream) [33]. In this study, we found in addition to the sites reported previously two additional sites were identified 9.6 kb and 14.6 kb upstream of the gene (Figure 2D). An additional example is the integrin alpha M chain (CD11b) gene, ITGAM. This gene is known to be important for the adherence of neutrophils and monocytes during differentiation. It contains one PU.1 binding site at the promoter as previously reported [12]. In our study we have identified another site 16 kb upstream of the gene, representing a potential enhancer region (Figure 2D). Such structures may typically represent PU.1-involved trans-cis interactions required for myeloid differentiation. In consistent with this notion, numerous other myeloid differentiation-required genes are also multi-targeted by PU.1, such as LMO2, BCL3, IL1B and IL12B (Figure 2D and Additional file 5: Figure S3).

To identify cis-elements that interact with PU.1 in the enriched binding regions, we conducted motif discovery analysis using several discovery tools, including (1) the de novo motif discovery methods AMD [34] and MEME [35], and (2) the prior-compiled PSFM-based motifs detection method termed MotifScan (see Materials and methods). Based on sequences corresponding to the top 500 most-enriched binding regions, two types of motifs appeared to be repeatedly observed in a significant manner (Figure 3A). One is contained in the database of TRANFAC and known as the canonical PU.1 consensus sequence of 5^′-AG(A/G)GGAAG-3^′ (left panel) and the other is found by de novo scanning, with a motif sequence of 5^′-(A/G)AAAG(A/G)GGAAGTG-3^′ (right panel and Additional file 6: Figure S4). This de novo motif covers the canonical one but identifies additional preferences including adenine at the -3 to -1 position, and thymine and guanine at the +1 and +2 positions. For convenience, we have named the motif including 5^′ adenines as “long motif” and the canonical one as “short motif”.

Next, we scanned the total binding regions of PU.1 using the above long and short motifs by MotifScan. As shown in Figure 3B, 37.1% of the binding regions contained one or more long motifs, and 46.3% of the binding regions contained only the short motif. The remaining 16.6% revealed neither long nor short motifs, which were likely due to undetected PU.1-binding motifs present in these regions, or due to the possibility that for these sites PU.1 does not directly bind to chromatin, but rather forms a complex through protein-protein interactions. Positional distribution analysis revealed that both motifs, especially the long one, appeared to reside near the center of the binding regions (Figure 3C). We then evaluated the binding affinities of the long and short motifs by comparing their enrichment levels. As shown in Figure 3D, the long motif exhibited a much higher mean tag density than the short one, particularly with respect to the tag density in the regions (-100 to +100) flanking the summit of peaks. Consistently, we found that the higher the enrichment levels, the more (less) the percentages of the long (short) motifs (Figure 3E). These observations suggest that the long motif exhibits higher binding affinity to PU.1 than the short one. Interestingly, sequence evolutionary conservation analysis showed that the binding sites with the short motif appeared to be much more conserved than the long motif-containing sites (Figure 3F). These results together suggest that the motif preference may correlate with the motif location in the genome, implicating that functional roles played by the two types of motifs can be different in general. Then, we examined the proportional distributions of the short or long motifs in promoter regions vs. non-promoter regions, respectively. As shown in Figure 3G, the percentage of short or long motifs (48.8% vs. 36.1%, outer circle) in non-promoter regions was equivalent to that for the total of PU.1 binding sites (46.3% vs. 37.1%, Figure 3B), whereas, that in the promoter regions with 63.3% short motifs and 14.5% long motifs (inner circle) was significantly different from that.

In sum, cis-elements that interact with PU.1 can be classified into short and long motif classes based on their sequence patterns. In promoter regions, PU.1-cis elements are predominantly represented by those in the class of short motif, which are highly conserved across species but with lower binding affinity to PU.1. In contrast, cis-elements in the class of long motif are relatively depleted from promoter regions whereas primarily present in non-promoter regions, representing PU.1 binding sites of high affinity but evolutionally less conserved.

We previously demonstrated that PML/RARα selectively targets PU.1 binding regions that harbored both PU.1 binding sites and RARE (retinoic acid response element) half (RAREh) sites [6]. In the present study, we found that cis-elements that interact with PU.1 could be distinguished by the sequence length and composition, evolutionary conservation, genomic distribution and binding affinity. A remaining question would be whether these motifs were functionally differentiated, e.g., in the genesis of APL. In an attempt to answer this question, we examined PU.1 binding sites in the genomic regions targeted by PML/RARα. First, ChIP-seq was performed using specific antibodies against PML and RARα respectively, and a total of 3,551 highly significant (FDR < 0.01) PML/RARα binding sites were identified (Additional file 7: Table S3). When the binding sites of PML/RARα and those of PU.1 were compared, over 53% (1,886/3,551) of the PML/RARα were also targeted by PU.1 (Figure 4A). When short and long elements of PU.1 motifs were respectively compared in the PU.1-specific binding sites, and PU.1 and PML/RARα overlapping binding sites (PU.1&PR), we observed dramatic differences (Figure 4B). For the PU.1&PR set, almost 70% of the sites were represented by the short PU.1 motif while less than 12% were typified by the long motif. In contrast, the proportion in the PU.1-specific set was 49.5% long and 34.7% short. These results indicated that the short motif elements are strongly correlated with the selective targeting of PU.1 binding sites by PML/RARα whereas the long motif elements are largely depleted from such a targeting. Although much remains to be elucidated, our results, at least, suggest that binding of PU.1 to short motif elements may offer this factor a selective preference to recruit PML/RARα. Previously, we have shown that PU.1 is able to direct the binding of PML/RARα to nearby RARE half (RAREh) sites at the level of promoters [6]. We thus extended this analysis to a whole genome scale and analyzed the enrichment of RAREh sites in the three subpopulations (PU.1-specific, PU.1&PR and PR-specific). As illustrated in Figure 4C, RAREh sites were significantly enriched in the PR-specific (Z-score = 22.3) and PU.1&PR (Z-score = 21.4) sets, whereas the RAREh sites were relatively depleted from the PU.1-specific set (Z-score = 3.35). This result further indicates that the selective binding of PML/RARα to PU.1 binding sites requires both RAREh and short consensus elements of PU.1.

Next, we wanted to know whether these PU.1-bound regions with distinct cis-elements could be co-bound or tethered differentially by other factors in addition to PML/RARα. For this analysis we took advantage of the published Encyclopedia of DNA elements (ENCODE) project data that includes the genomic regions bound by 119 human transcription factors involving 72 different cell types [36, 37]. We compiled three different PU.1-bound region sets, including the PU.1-specific with the long motif (Long), the PU.1-specific with the short motif (Short) and the PU.1&PR, and carried out overlapping analysis with 1328 ChIP-seq data sets downloaded from the UCSC ENCODE data center. As shown in Figure 4D and 4E, most of the ENCODE TFs showed significantly less preference to the Long set than the Short set and the PU.1&PR set (paired t-test p<1e-303 and p<1e-296, respectively). In particular, factors, such as MAX (21.3% vs. 37.7%) and MYC (17.8% vs. 34.4%) in NB4 cells, JUND in K562 cells (13.4% vs. 26.0%) and CTCF in HCPE (12.3% vs. 27.3%) and HBMEC (12.9% vs. 25.9%) cells showed the most distinct overlapping percentage between the Long and Short sets. In contrast, the Long set was significantly more covered by 9 ENCODE PU.1 data sets involving 4 cell types (GM12878, GM12891, HL60 and K562) than the Short and PU.1&PR sets (paired t-test p<1e-6 and p<0.0002, respectively; Figure 4F). The result suggested that the long-motif-containing sites kept more PU.1 binding stability among different cell types, consistent with their higher binding affinity to PU.1 revealed above, while the short-motif-containing sites could recruit more TFs to regulate the corresponding gene expression.

Furthermore, we wanted to know whether genes regulated by different PU.1 cis-elements or binding factors were functionally differentiated. Thus, we conducted Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses respectively on the three gene sets (Long, Short and PU.1&PR). As shown in Figure 4G, the most striking GO terms revealed in the gene set targeted by both PU.1 and PML/RARα (PU.1&PR) were highlighted by genes involved in oncogenesis and hematopoiesis. The former was represented by JUNB, BCL2, MLLT10, FES, FOS and MCL1, and the latter was represented by RUNX1, CEBPA and STAT3. Similarly, the most significant KEGG pathways revealed in this gene set were represented by those involved in acute myeloid leukemia, including SPI1, RARA, KRAS and so on (Figure 4H and Additional file 8: Figure S5). These results provide additional evidence that genes targeted by both PU.1 and PML/RARα are indispensable for the normal hematopoiesis, and are crucial for leukemogenesis (Figure 4G and 4H). Besides, we also noticed that the genes belong to the Long set specifically evolved in several biological processes and functional pathways, such as developmental processes and adherens junction, respectively. The Short gene set, however, showed little particularly functional enrichment, implying PU.1 may involve more extensive biological functions rather than some specific ones through the different dynamic combinations with other factors.

NB4 cells were cultured in RPMI 1640 medium supplemented with 10% fetal bovine serum (FBS). HEK 293T cells were cultured in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% FBS, in a humidified atmosphere with 5% CO₂ at 37°C.

Promoter and enhancer regions harboring PU.1 motifs were cloned into the pGL3-basic and pGL3-promoter vector (Promega, Madison, WI), respectively. The primers used for the plasmids constructs are listed in Additional file 9: Table S4. The renilla luciferase plasmid pRL-SV40 (Promega, Madison, WI) was used as control for transfection efficiency. The expression plasmid was pCMV4-PU.1. HEK 293T cells were transiently transfected using Lipofectamine 2000 (Invitrogen, Carlsbad, CA). Transfected cells were cultured for 48 hours and then assayed for luciferase activity using Dual-Luciferase Reporter Assay System reagents (Promega, Madison, WI).

ChIP-qPCR was performed using Power SYBR® Green PCR Master Mix (Applied Biosystems, Foster City, CA) and ABI Prism 7900HT detection system (Applied Biosystems, Foster City, CA). The fold enrichment of the tested binding regions over the input DNA was estimated as previously described [45]. The primers used for ChIP-qPCR are listed in Additional file 10: Table S5.

ChIP was performed using specific antibodies according to the Affymetrix protocol as described previously [40]. ChIPed and Input DNA were sequenced with Illumina Genome Analyzer II. The 35 bp reads (or tags) were aligned (mapped) to the unmasked human reference genome (NCBI v36, hg18) using the Eland application (Illumina) allowing two mismatches. Only uniquely mapped reads were retained to further analyses. Next, MACS [16] algorithm was used to identify PU.1 binding regions. For the visualization of the enrichment level of transcription factor’s binding sites, we calculated the tag density of each ChIP-seq sample with 500 bp window, aligned it to the same coordinate and visualized them in a bar plot using the IGB (Affymetrix) program.

We used the RefSeq Genes’ database from UCSC to map and annotate the peak regions. For each peak region, we first searched the nearest RefSeq gene in both directions unless no gene was found within 50 kb. For a peak region lying within a gene, we classified it to the proximal promoter (-2 kb upstream to 1kb downstream to TSS), gene body (1 kb downstream of the TSS to the transcription end site (TES)), upstream enhancer (between at most -50 kb and -2 kb upstream to the TSS), or downstream enhancer (from TES to at most 50 kb downstream). Otherwise, we marked these peaks as distal intergenic region (>50 kb from a RefSeq gene). To avoid multiple genomic region type annotation of one peak, we uniquely mapped the peaks to genomic region type following the priority rule: proximal promoter > gene body > upstream enhancer > downstream enhancer > distal intergenic region.

The enriched peaks were first uniquely mapped to certain genomic regions according to the above peak mapping and annotation criteria. Then, the regions were aligned at their summits from 5^′ to 3^′ in accordance with the orientation of the corresponding genes (if a peak belongs to the distal intergenic region, we arbitrarily assume that the peak is on the positive strand) and uniformly expanded to 3,000bp in each direction, and phastCons scores were retrieved from UCSC genome browser (http://​genome.​ucsc.​edu) and averaged at each position.

A position-specific frequency matrix (PSFM) similarity based motif analysis algorithm, named MotifScan, was used to analyze the enrichment of the known motifs statistically on the ChIP-seq data. Fold enrichment and Z-score were used to assess the significance of motif enrichment. MotifScan was performed as follow:

For a specific motif, the sequence with a similarity higher than the threshold was marked as a matched sequence. Suppose that the total number of matched sequences was X in a data set with total effective sequence length L bp (repeat masked). Then the random variable X followed the binomial distribution with parameters n and p, written as X ∼ B(n, p), where n ≈ L, and p was the possibility that each bp contained a matched sequence. In this case, n was very large generally. Meanwhile, n*p and n*(1-p) were large enough. Thus, to simplify the calculation, we used the normal distribution X ∼ N(np, np(1 − p)) to excellently approximate the binomial distribution. Subsequently, we further simplified the 1–p as 1, since the p was small generally. Therefore, the expected value (E) and the variance (σ²) of X would be the same.

Next, the total number of matched sequences was counted in a background data set and a sample data set, respectively. Suppose that the total number of matched sequences is N _s in a sample data set with total effective sequence length L _s (bp), and N _c in a background data set with total effective sequence L _c (bp). We calculated the expected number of matched sequences in the sample data set as E = N c × L s L c Followed the above simplified formula, the fold enrichment and Z-score were calculated as F = N s E and Z = N s − E E, respectively.

For a specific motif, if a peak region contained at least one matched sequence, then we marked this peak as a motif-containing one. The percentage of certain motif-containing peaks was calculated as the number of motif-containing ones divided by all the number of peaks. The similarity was calculated as the same as in MotifScan algorithm. In this study, to enhance the accuracy and specificity, the similarity threshold for the “long” PU.1 motif was set at 0.8 and that for the “short” PU.1 motif and RARE half was set at 0.9. We first excluded all the long PU.1 motif-containing peaks when calculating the only short PU.1 motif-containing ones, as the short PU.1 motif was nearly covered by the long one.