The DEK oncoprotein binds to highly and ubiquitously expressed genes with a dual role in their transcriptional regulation

For a genome-wide analysis of DEK binding, we performed chromatin immunoprecipitation of DEK in the U937 cell line followed by high-throughput sequencing (ChIP-seq). We identified 4581 peaks of DEK binding, determined as overlapping peaks in two independent precipitations (Additional file 1: Table S1). The specificity of the immunoprecipitating antibody was verified by Western blot (Figure 1D). To determine the dispersion pattern of DEK throughout the genome, we calculated the distance from the middle of each DEK peak to the nearest transcription start site (TSS). The result shows a strong accumulation of DEK binding around the TSS (Figure 1A-B). This was confirmed by analysis with the Nebula software, showing that the highest enrichment of binding occurs in promoter regions (Figure 1C).

To characterize the genes bound by DEK, we stratified them based on their expression levels. All the genes in the genome were assembled into groups of one thousand, from the most to the least expressed genes based on previously established expression levels in the U937 cell line, as determined by CAGE analysis [32]. For each group, we then calculated the proportion of genes that were bound by DEK. To exclude DEK binding which is less likely to be relevant for gene regulation, we selected the genes to which DEK binds within 1 kb from the transcription start site but similar results were obtained using other or no boundaries. The result clearly demonstrates that DEK mainly binds to highly transcribed genes (Figure 2A). Furthermore, we divided all the DEK-bound genes into four groups based on their expression; no (CAGE 0), low (CAGE 0–10), intermediate (CAGE 10–100) and high (CAGE >100) expression. We then calculated the distances from each DEK binding site to the nearest transcription start site. Interestingly, the genes that were not expressed lacked the characteristic accumulation of DEK at the transcription start sites (Figure 2B). And for the expressed genes, higher expression correlated with higher accumulation of DEK (Figure 2C-E). The correlation between DEK binding and gene transcription was further substantiated by comparing the DEK binding pattern to those of previously characterized factors. We constructed an algorithm that scored and ranked the binding patterns from all 2642 ChIP-seq experiments in the Encode database based on their similarity with the DEK binding pattern (Additional file 2: Table S2). Strikingly, many of the most similar binding patterns represented ChIP-seq analyses of RNA polymerase II (POL2) binding, indicating active gene transcription. This was visualized by ranking the ChIP-seq experiments in Encode from most to least similar to DEK, with POL2 experiments represented by black bars (Figure 3A). The analysis clearly shows accumulation of POL2 experiments among the patterns most similar to that of DEK, with half of the 100 most similar binding patterns belonging to POL2. The same result was obtained when the analysis was conducted only with Encode binding patterns established in myeloid cells (Additional file 3: Table S3). In all four myeloid datasets, the POL2 binding patterns were among the very most similar to the DEK binding pattern (Additional file 4: Table S4). The Encode database contains experiments from a wide range of cells and tissues. Thus, the striking similarity between DEK and POL2 binding across the entirety of the database, as seen in Figure 3A, shows that DEK binding is associated with genes that are commonly transcribed across different cell types rather than tissue- or lineage-specific genes. To confirm this notion, we constructed figures as the one in Figure 3A for each transcription factor in the Encode database. We then calculated the similarity between the binding pattern of each factor and the complete collection of POL2 experiments, as the rank of the median POL2 experiment. Finally, we compared these scores with that of DEK. Out of the 1172 analyzed binding patterns, only 39 had a higher correlation with overall POL2 binding than DEK did (Additional file 5: Table S5). With few exceptions, the precipitated factor in these cases was POL2 itself. DEK even ranked higher than 66% of the POL2 binding patterns. This confirms that DEK binds to genes that are not only highly expressed but also ubiquitously expressed across different cell types.

Given the binding of DEK to highly expressed genes and the implication of DEK in epigenetic regulation, we wanted to characterize the epigenetic landscape of DEK binding, To this end, we compared the DEK binding pattern with the Encode experiments termed OpenChromSynth, which indicate regions of open chromatin, as determined by a combination of ChIP-seq, FAIRE-seq and DNAse hypersensitivity analysis. The result shows high similarity between these genomic regions and the DEK binding pattern (Figure 3B), suggesting extensive overlap between DEK binding and regions of open chromatin. The same result was obtained when the analysis was restricted to the myeloid datasets (data not shown). To determine the histone modifications associated with DEK binding, we calculated the similarity between the DEK binding pattern and those of the histone modifications analyzed in the BroadHistone K562 dataset in Encode (Figure 4). In concordance with binding to highly transcribed genes, DEK showed the highest degree of overlap with the active TSS marker H3K4me³, the active enhancer marker H3K27ac and the active promoter markers H3K4me² and H3K9ac [33‐36]. Conversely, the two repressive histone modifications H3K27me³ and H3K9me³ (grey bars) were the ones with the least overlap [37]. The association between DEK and active chromatin was further demonstrated by performing a DNA methylation array to calculate the methylation grade for each gene. The methylation of the genes bound by DEK was then compared to the global gene methylation. The result shows markedly lower methylation of the genes to which DEK binds (Figure 5). Thus, consistent with the binding to the transcription start sites of highly and commonly transcribed genes, we also find that DEK binds to regions of open chromatin with low DNA methylation and which are rich in activating histone marks.

The results above show that DEK binds to genes commonly expressed in different cell types and tissues. This was further confirmed by gene ontology analysis of the genes to which DEK binds within 1000 bp of the transcription start site. The significantly enriched categories mainly represent basic metabolic processes that are shared by all cells (Table 1). Several are however frequently deregulated in cancer and essential for cancer cell proliferation, such as protein synthesis and cell cycle regulation. Notably, the category with the highest enrichment of DEK-bound genes is “Nucleosome assembly”, a process in which DEK itself is directly and functionally implicated [38]. This suggests that DEK could contribute to epigenetic regulation not only directly but also by binding and regulating other genes involved in the process. Gene ontology analysis of the genes deregulated by knockdown of DEK in U937 cells by shRNA shows that many of the genes whose expression is affected by DEK are involved in the same processes, such as cell cycle and chromatin regulation (Additional file 6: Table S6). Since no binding motif has been defined for the DEK protein, we analyzed the DEK-bound sequences for enriched motifs (Table 2). The identified motif with the highest degree of significance was found to be a binding motif for the transcription factor PU.1, a key regulator of gene expression during myeloid cell development and highly expressed in these cells. The other known motifs belong to the transcription factors SP1, RUNX1 and USF1, all of which are also expressed in these cells. Although we also identified several motifs not corresponding to known transcription factors, it is unlikely that any of these represent genuine DEK motifs, since none of them predict DEK binding as well as the PU.1 motif.

The role of DEK in gene regulation is far from understood. Previous studies have determined both activating and repressive effects on the expression of single genes. To study the effect on a genome-wide scale, we combined the results of the ChIP-seq analysis with gene expression microarray analysis following knockdown of DEK by shRNA. Based on the microarray analysis, we characterized the DEK-bound genes as either upregulated, downregulated or unaffected by the knock-down of DEK (Additional file 7: Table S7). To only include the DEK binding most likely to influence gene expression, the analysis was limited to genes where DEK binds less than 1000 bp either upstream or downstream from the transcription start site. Compared to the complete genome, significantly more genes were found to be either upregulated or downregulated by DEK depletion than unaffected. The same result was obtained when either fold change (Figure 6A, E) or statistical significance (Figure 6B) were used as the measure of change in gene expression. The same pattern was observed when DEK was knocked down in primary CD34⁺ cells isolated from human umbilical cord blood (Figure 6C-D). That the correlation is not as strong for the primary cells is expected, given the discrepancy in gene expression between these cells and the U937 cell line in which the ChIP-seq analysis was performed. These findings emphasize the role of DEK as a gene regulatory protein and indicate that DEK serves as both an activator and a repressor of transcription in the same cell under the same conditions. To validate the association of DEK binding with gene expression, we performed a network analysis of a collection of publicly available microarray datasets from patients with acute myeloid leukemia. For all of the 1246 analyzed transcription factors, we calculated the number of genes whose expression correlated with that of each transcription factor (Table 3). Interestingly, DEK was found to be the factor with the third highest number of correlated genes. The analysis was then repeated using microarray data from the Microarray Innovations in Leukemia (MILE) study [39]. In this analysis, DEK ranked in the 11^th percentile and out of the 30 factors with the most connections in the original dataset, DEK was the one with the most connections in the MILE dataset. Taken together, the network analyses underscore the importance of DEK in gene regulation.

The U937 cell line (ATCC, Manassas, VA, USA) was cultured in RPMI 1640 medium (Life Technologies, Carlsbad, CA, USA) supplemented with 10% fetal bovine serum (Life Technologies). Primary CD34⁺ cells were obtained from human umbilical cord blood collected at Skåne University Hospital. The mononuclear cell population was isolated by separation on Lymphoprep (Axis-Shield PoC AS, Oslo, Norway) and CD34⁺ cells were subsequently selected using the Indirect CD34 MicroBead Kit (Miltenyi Biotec, Bergisch Gladbach, Germany). The cells were grown in StemSpan SFEM medium (Stemcell Technologies, Vancouver, Canada) supplemented by 20% fetal bovine serum and the CC100 cytokine cocktail (Stemcell Technologies).

Chromatin immunoprecipitation was performed on U937 cells using the Magna-ChIP A/G Chromatin Immunoprecipitation Kit (Merck Millipore, Billerica, MA, USA) after crosslinking with 1% formaldehyde for 15 min. Chromatin shearing was achieved by sonication in a Bioruptor UCD-200 (Diagenode, Liège, Belgium). Immunoprecipitation was performed with 10 μg of DEK antibody (Abcam, Cambridge, UK; product code ab74975) per million cells. Immunoprecipitated DNA was sequenced by the Science for Life Laboratory in Stockholm, Sweden with an Illumina HiSeq 2000 as paired-end reads to 100 bp with a minimum of 18 million reads per sample. The Illumina OLB v1.9 was used for base conversion, the Bowtie 2 software [44] was used for the alignment of reads to the hg19 reference genome and peak calling was performed with the MACS software v1.4.1 [45]. Default parameters were used throughout the process. Two independent ChIP-seq experiments were performed under identical conditions and the peaks found in both experiments were used for all subsequent analyses. Non-precipitated chromatin was used as negative control. Validation was performed by real-time PCR analysis of DNA immunoprecipitated with the DEK antibody, showing enrichment of DNA corresponding to the predicted binding sites in the S100A9 (fold enrichment 2.4) and VIM (fold enrichment 1.8) genes and a lack of enrichment of DNA corresponding to the IRF8 gene (fold enrichment -3.1), which was determined by the ChIP-seq analysis to not be bound by DEK. The raw data from the ChIP-seq analysis is available through the Gene Expression Omnibus data repository, with the accession number “GSE60692”.

The “Genomic annotation of ChIP-seq peaks” tool in the Nebula software package [46] was used with default parameters to find the closest gene to each DEK peak and calculate the distance from the transcription start site to the middle of the peak. The same tool was used to determine the genomic elements to which DEK binds, using default settings where promoters are designated as the 2.000 bp upstream of the transcription start site and enhancers are designated as the 30.000 bp upstream of the TSS. As control, random genomic sequences were generated by a random draw with replacement from a square-distribution across all genomic positions.

Absolute gene expression levels for all genes in the U937 cell line were determined by Cap Analysis of Gene Expression (CAGE) in a study by the FANTOM5 consortium [32]. All genes were ordered by absolute expression and divided into bins of one thousand, for which we calculated the number of genes where DEK binds within 1000 bp from the transcription start site. For further analysis, all DEK-bound genes were divided into four categories according to their expression; no (0), low (0–10), intermediate (10–100) or high (>100) expression. For each category, the distances to the nearest transcription start sites were calculated as described above.

The determined DEK binding pattern was compared to the binding patterns previously established for transcription factors and histone modifications as well as the genome-wide distributions of characteristics such as DNAse hypersensitivity. To this end, we constructed an algorithm which analyzed all of the 2642 such patterns in the Encode database and scored them based on their similarity with the binding pattern of the DEK protein. For each Encode binding pattern, the weight of each base was calculated based on the number of sequence tags at that position in the genome. The raw correlation between this sequence tag distribution and that of DEK was then calculated as the sum of the products of the weights of each matching base, corrected for track length using the inverse square of the total weight of each track. Formally,

where weight1_i and weight2_i are the score assigned basepair i in the two tracks compared, respectively, and n_basepairs is the total number of basepairs in the hg19 assembly of the human genome. DEK binding was correlated with open chromatin status by comparison of the DEK binding pattern with the Encode tracks designated OpenChromSynth, each containing a synthesis of results from ChIP-seq, FAIRE-seq and DNAse hypersensitivity analysis in the examined cell type. The binding patterns of different histone marks in the K562 cell line were obtained from the Broad Histone dataset in Encode. Statistical testing of the enrichment of tracks representing either RNA polymerase II or OpenChromSynth among the most similar binding patterns to that of DEK was performed using the RenderCat software, as previously described [47].

Microarray-based DNA methylation analysis was performed with the Infinium HumanMethylation27 BeadChip (Illumina) by the BEA core facility at Karolinska Institute. Genomic DNA from U937 cells was extracted using the QIAamp DNA mini kit (Qiagen) and bisulfite converted using the Zymo EZ DNA methylation Kit (Zymo Research, Irvine, CA). Subsequently, the DNA was subjected to the Illumina Infinium HD Methylation assay including whole genome amplification and enzymatic fragmentation before hybridization to the BeadChip. Arrays were scanned and the signals processed in Genome Studio module 1.8. The methylation grade for each gene was calculated as the average of the corresponding probes. The methylation of the genes to which DEK binds within 1000 bp of the transcription start site was then compared to the methylation of the entire genome. The raw data from the DNA methylation analysis is available through the Gene Expression Omnibus data repository, with the accession number “GSE60734”.

Two shRNAs in H1 lentiviral vectors targeting the DEK transcript were a kind gift from Dr David Markovitz [2]. Lentiviral particles were harvested after calcium phosphate transfection of 293 T cells (ATCC) with the respective shRNA constructs, gag-pol and the RD114 envelope gene. For lentiviral transduction, non-tissue culture coated plates were coated with retronectin (Takara, Otsu, Japan) and blocked with 2% bovine serum albumin (Sigma-Aldrich, St. Louis, MO, USA) for 30 min at room temperature. Subsequently, virus-containing medium was added and the plates were centrifuged at 1000 × g for 60 min at 4°C before the cells were added and incubated at 37°C for 48 h, after which they were sorted by FACS based on the expression of the GFP marker. Transduction efficiencies were similar for all constructs and consistently above 40%. The efficiency of the knockdown was verified by Western blot of cell lysates obtained 4 days (U937 cells) or 10 days (primary cells) after sorting, with a primary DEK antibody (BD Transduction Laboratories, San Jose, CA, USA). The raw data from the gene expression analysis is available through the Gene Expression Omnibus data repository, with the accession number “GSE60734”.

RNA was isolated with the RNeasy Mini Kit (Qiagen, Hilden, Germany) from cells harvested 4 days (U937) or 10 days (primary cells) after sorting. Microarray analysis was performed in triplicates with the HumanHT-12 v4 Expression BeadChip (Illumina, San Diego, CA, USA) by the Swegene Centre for Integrative Biology at Lund University (SCIBLU). Designated as upregulated or downregulated were the genes found to be among the thousand most strongly upregulated or downregulated based on either the average fold change or the combined p value from statistical testing of the two shRNA constructs by the student’s t test. Subsequently, the number of genes to which DEK binds within 1000 bp from the transcription start site was calculated for each category. The gene expression microarray was validated by quantitative real-time PCR analysis of two genes found to be upregulated (CSF2RA; fold change 1.3, p = 0.004 and CSF3R; fold change 1.3, p = 0.04) and two genes found to be downregulated (CHMP2B; fold change −1.3, p = 0.03 and ICAM2; fold change −1.4, p = 0.002) upon knockdown of DEK in the U937 cell line. For this analysis, RNA was reverse-transcribed to cDNA with the High Capacity cDNA Reverse Transcription Kit (Life Technologies) and real-time PCR was performed using the TaqMan Gene Expression Assay (Life Technologies) and the StepOne Plus Real-Time PCR System (Life Technologies).

Gene ontology analysis of the genes bound by DEK was performed with the GO::TermFinder software [48], using gene ontology associations based on the UniProt reference proteome. The analysis was performed on the genes to which DEK binds within 1000 bp from the transcription start site, with the default p value threshold of 0.01. Fold enrichment was calculated as the percentage of the DEK-bound genes associated with a certain term divided by the percentage of the total genome associated with the same term. The result was filtered for redundant terms and terms containing more than 5000 genes. Gene ontology analysis of the genes deregulated by the knockdown of DEK was performed by Gene Set Enrichment Analysis (GSEA), using a p value threshold of 0.01 and a false discovery rate threshold of 0.05 [49].

Motif analysis was performed with the Discriminative Regular Expression Motif Elicitation (DREME) software [50], using as input the sequences from 150 bp upstream to 150 bp downstream of the middle of all DEK peaks. The p value threshold was set to 0.01. Identified motifs were subsequently matched with known motifs using the TOMTOM software [51] with a false discovery rate threshold of 5%.

The genome-wide AML network of gene expression was constructed by LASSO regression modeling of gene expression correlations in 3013 samples obtained by merging samples from studies based on the Affymetrix HG-U133 Plus 2.0 GeneChips (GPL570) platform (see Additional file 8 Methods for a complete list of datasets). The MILE network was constructed in the same manner but instead based on the results from stage one of the Microarray Innovations in Leukemia study [39]. To reduce the false discovery rate, the correlation threshold of the network models was set to a level at which they did not find any correlations in randomized data. The list of transcription factors used in the analysis was manually curated after its original development as part of the Differentiation Map project [52].