Contributors:Leighton J. Core, Joshua J. Waterfall, Daniel A. Gilchrist, David C. Fargo, Hojoong Kwak, Karen Adelman, John T. Lis
UCSC Browser Screenshot of Sarkosyl Dependent and NELF RNAi GRO-Seq Experiments, Related to Figure 2
TSS-RNA reads (Nechaev et al., 2010) marking TSSs are in dark blue for the plus and minus strand (read/base/10ˆ6 reads). GRO-seq reads (reads/base/10ˆ6 reads) aligning to the plus strand are shown in red; minus strand in blue. ChIP-seq for total Pol II (α-Rpb3) is shown in green (reads/25bp bin), and gene annotations are shown at the bottom in blue. The arrowheads depict TSSs and the ∗ denotes a TSS that is reannoated with our data sets.
... Pol II at Promoters Is Predominantly Engaged and Competent for Elongation
(A) Representative browser shot showing Pol II Chip-seq (green) and GRO-seq (red) with y axis in reads/bp/10exp6. The regions used for calculating the engaged and competent fraction (ECF) at promoters are indicated below.
(B) Schematic explaining the workflow used to calculate the ECF for Pol II at promoters.
(C) Histogram showing the distribution of ECF values for significantly bound promoters (n = 3,168). The vertical lines represent a 50% (black), the average (red), and the Hsp70 (green) ECFs. Promoters with the lowest ECFs are highlighted in purple.
(D) Boxplots showing Pol II ChIP-seq levels at promoters with different ranges of ECF. Promoters with the lowest (purple) and the highest (dark red) ECF values have less Pol II bound at promoters in ChIP-seq experiments than promoters with less extreme ECF values (middle 20% shown), suggesting that the ChIP and GRO discrepancies here could be due to experimental noise.
The box spans the first quartile (Q1, bottom) to third quartile (Q3, top), the horizontal line in the box represents the median, and the whiskers extend as follows: (Q1 or Q3 + 1.5 )∗(Q3-Q1). See also Figures S6 and S7.
... RNA Polymerase Distribution on mRNA-Encoding Genes Using GRO-Seq
(A) A representative view of GRO-seq data from S2 cells in the UCSC genome browser (Kent et al., 2002). GRO-seq reads (reads/base) aligning to the plus strand are shown in red; minus strand in blue. ChIP-seq for total Pol II (α-Rpb3) is shown in green (reads/25 bp bin), and gene annotations are shown at the bottom in blue.
(B) GRO-seq data aligned to transcription start sites (TSSs). For all genes, reads aligning to the sense strand of the gene are in red; antisense strand in blue. For nonbidirectional genes (head-to-head promoters within 1 kb removed), reads aligning to the sense strand of the gene are in green; antisense strand in orange.
(C) Comparison of directionality of Drosophila and human promoters. The distribution of the ratios of sense and antisense reads around promoters (log2) is plotted for active promoters (>25 reads) in IMR90 cells (green) and Drosophila S2 cells (blue). How different types of directionality of transcription from promoters are reflected in the ratio are indicated in italicized lettering.
(D) GRO-seq profiles from ±1.5 kb relative to TSS are shown for all human promoters (green, sense; orange, antisense) or human promoters that contain a TATA box (red, sense; blue, antisense).
(E) GRO-seq data aligned to gene end for all genes (red, sense; blue, antisense), and after convergent genes within 1.5 kb are removed (green, sense; orange, antisense).
See also Figures S1 and S2.
... Supporting Genomic Data at Enhancers in This Study, Related to Figure 4
(A) GRO-seq data at putative human enhancers (n = 34,915). GRO-seq data is from IMR90 cells (Core et al., 2008). Data was compiled relative to the center of DHS sites.
(B and C) Pol II ChIP-seq and NELF ChIP-chip data (Gilchrist et al., 2010), respectively, around Drosophila putative enhancers.
... Comparison between Assays that Detect Polymerase at Promoters and in Genes, Related to Figure 5
(A–C) Shown are scatter-plots comparing amount of sequencing reads between (A) GRO-seq and Pol II ChIP-seq, (B) small-RNA-seq to ChIP-seq, and (C) GRO-seq to small-RNA-seq at promoters. All unique promoters are shown in black (n = 12,541); promoters called Pol II-bound by ChIP-seq (n = 3,168) in red. rho is Spearman's correlation coefficient between the two data sets.
(D) Normalization between ChIP-seq and GRO-seq data sets through fitting of signal within gene bodies. Plotting of the signal for each assay within all genes (black) shows a poor correlation (rho = 0.54), whereas plotting the signal for each after selecting genes that are highly active gives (red) an excellent correlation between data sets (rho = 0.87). Highly active genes are classified as those with the top 10% of Ser2P ChIP (n = 1874) signal within the gene (mark of active polymerases). The fit of the gene data for highly active genes is shown in green. This equation is used for calculating the engaged, competent fraction of polymerase at promoters.
Contributors:David A. Orlando, Mei Wei Chen, Victoria E. Brown, Snehakumari Solanki, Yoon J. Choi, Eric R. Olson, Christian C. Fritz, James E. Bradner, Matthew G. Guenther
Normalization and Interpretation of ChIP-Seq Data
(A) Schematic representation of a typical ChIP-seq data workflow. Interrogation of a human epigenome (Blue circles, nucleosomes) with a full complement of histone modification (red circles, top) versus an epigenome with a half complement of histone modification (red circles, bottom). ChIP, sequencing, and mapping using reads per million (RPM) reveals ChIP-seq peaks (blue). A comparison of the peaks as a percentage of the total reads reveals little difference.
(B) Schematic representation of a ChIP-seq data workflow with reference genome normalization. Interrogation of a human epigenome (Blue circles, nucleosomes) with a full complement of histone modification (red circles, top) versus an epigenome with a half complement of histone modification (red circles, bottom). A fixed amount of reference epigenome (orange, nucleosomes; red, histone modifications) is added to human cells in each condition. After ChIP, sequencing, and mapping, the ChIP sequence reads are normalized to the percentage of reference genome reads in the sample (reference-adjusted RPM [RRPM]). A comparison of ChIP-seq signals using normalized reads reveals a 50% difference between peaks. This method is called ChIP with reference exogenous genome (ChIP-Rx).
... ChIP-Rx Reveals Quantitative Epigenome Changes
(A and B) Percentage of reads aligning to either test (human, blue) or Drosophila (reference, orange) genomes after H3K79me2 ChIP-Rx (A) or H3K4me3 ChIP-Rx (B). Samples containing 0%, 25%, 50%, 75%, or 100% EPZ5676 treated Jurkat cells were used as defined in Figure 2B.
(C and D) Sequenced reads from H3K79me2 (C) and H3K4me3 (D) immunoprecipitations at the RPL13A gene locus in traditional reads per million (RPM,top) or reference-adjusted reads per million (RRPM, bottom; see Experimental Procedures). Color indicates the percentage of sample treated with EPZ5676. The gene model is shown below the track.
(E) Meta-gene profile of H3K79me2-occupied genes in Jurkat cells. Meta-gene profiles were produced with traditional RPM (left) or RRPM (right). Color indicates the percentage of Jurkat cell sample treated with EPZ5676 as in Figure 2B. Region −5 to +10 kb around the transcription start site (TSS) is shown. Meta-gene profile was derived from top 5,000 protein-coding genes as defined by total H3K79me2 signal in the 0% treated (untreated with EPZ5676) sample. A meta-gene profile representing all genes is shown in Figure S3.
(F) Meta-gene profile of H3K4me3-occupied genes in Jurkat cells. Meta-gene profiles were produced with traditional RPM (left) or RRPM (right). Color indicates the percentage of Jurkat cell sample treated with EPZ5676 as in Figure 2B. Region −5 to +10 kb around the transcription start site (TSS) is shown. Meta-gene profile was derived from top 5,000 protein-coding genes as defined by total H3K4me3 signal in the 0% treated (untreated with EPZ5676) sample. A meta-gene profile representing all genes is shown in Figure S3.
(G and H) Line graphs display the observed fold-change difference in average meta-gene signal across the −5 to +10 kb window around the TSS for each H3K79me2 (G) or H3K4me3 (H) ChIP sample (x axis) relative to the signal from the 0% treated population using traditional (gray) or reference (black) normalization.
See also Figures S2–S4 and Table S2.
... ChIP-Rx Reveals Epigenomic Alterations in Disease Cells that Respond to Drug Treatment
(A) Western blot showing the levels of H3K79me2 in MV4;11 cells after treatment for 4 days with increasing concentrations of EPZ5676.
(B) Percentage of H3K79me2 ChIP-seq reads aligning to either test (human, blue) or Drosophila (reference, orange) genomes after H3K79me2 ChIP-Rx from MV4;11 cells treated as in (A).
(C) Sequenced reads from H3K79me2 immunoprecipitations at the REXO1 gene locus in standard RPM (top) or RRPM (bottom) (see Experimental Procedures). Color indicates the concentration of EPZ5676 given to each sample. The gene model is shown below the track.
(D) Meta-gene profile of H3K79me2-occupied genes in MV4;11 cells. Meta-gene profiles were produced with traditional Reads Per Million (RPM, left) or Reference-adjusted Reads Per Million (RRPM, right). Color indicates the concentration of EPZ5676 used in each sample. The region −5 kb to +10 kb around the TSS is shown. Meta-gene profile was derived from top 5,000 protein-coding genes as defined by total H3K79me2 signal in the 0nM treated (untreated with EPZ5676) sample. A meta-gene profile representing all genes is shown in Figure S3.
(E) Line graph displays the observed fold-change difference in average meta-gene signal across the −5 to +10 kb window around the TSS for each H3K79me2 ChIP sample (x axis) relative to the signal from the 0 nM treated population using standard (gray) or reference (black) normalization.
(F) Box plots display the distribution of the observed fold change of H3K79me2 signal −5 kb to +10 kb around the TSS of all genes between the 0 nM and 5 nM treated samples (blue, MV4;11; green, Jurkat) for all genes using traditional (left) or reference-adjusted (right) normalization (see the Supplemental Experimental Procedures).
See also Figures S3 and S5 and Table S2.
... Experimental Design of Differential H3K79me2 Detection
(A) Schematic representation of differential H3K79me2 detection and normalization strategies. Two populations of cells were produced: a human epigenome (blue nucleosomes) with a full complement of H3K79me2 (red circles, top left) and a human epigenome (blue nucleosomes) with depleted H3K79me2 due to EPZ5676 exposure (top right). These cells were mixed in defined proportions in order to allow a dilution of total genomic histone modification (dark red to pink). Cell mixtures were subjected to ChIP-seq in the presence of the reference Drosophila epigenome (orange). ChIP-seq signals were calculated based on traditional or Drosophila-reference-normalized methods. See also Figure S1.
(B) Western blot validation of H3K79me2 depletion in Jurkat cells. Mixtures of 0%–100% EPZ5676-treated cells (0:100; 25:75; 50:50, 75:25; 100:0 proportions of [DMSO-treated:EPZ5676-treated] cells) were measured by immunoblot (IB) for the presence of H3K79me2, H3K4me3, or total histone H3 (loading control). Treated cells were exposed to 20 μM EPZ5676 for 4 days.
See also Table S1.
Contributors:Rajprasad Loganathan, Joslynn S. Lee, Michael B. Wells, Elizabeth Grevengoed, Matthew Slattery, Deborah J. Andrew
Results from DAVID clustering analysis of GO terms for genes activated by Ribbon based on microarray data and bound by Ribbon in the salivary gland based on ChIP-Seq.
... ChIP-seq analysis identifies Rib binding sites in salivary gland cells. (A) Schematic outline of the experimental approach to identify SG-specific Rib binding sites. ChIP-seq datasets were obtained from samples using two different GAL4 constructs to drive expression of UAS-rib-gfp in the SG. The overlap of binding events observed with both drivers enriches for SG-specific Rib binding. (B) Rescue of the SG phenotype in the rib1/ribP7 mutant background with fkh-Gal4::UAS-rib-GFP verified the functionality of UAS-rib-gfp construct used in the ChIP-seq experiments. (C) Tissue expression of fkh-GAL4 and sage-GAL4 drivers spanning the stages used for the ChIP-seq analyses. Arrowheads indicate the SG at different developmental stages. (D) SG-enriched ChIP-seq signals correspond to Rib binding events in the vicinity of two Rib target genes – Hsp70Ba and Obp99b.
... Drosophila... Rib SG binding sites overlap with genes expressed in the SG and with genes whose expression changes in rib mutants based on microarray analysis. (A) Venn diagram representing the overlap of genes from the ChIP-seq (494 genes bound by Rib in the SG), microarray (774 genes activated by Rib in the whole embryo) and BDGP gene expression database (434 SG-enriched gene expression). (B) Venn diagram representing the overlap of genes from the ChIP-seq (494 genes bound by Rib in the SG), microarray (1176 genes repressed by Rib in the whole embryo) and BDGP gene expression database (434 SG-enriched gene expression). (C) In situ hybridization analysis of SG genes activated by Rib and with nearby Rib binding sites in rib1/ribP7 mutant and heterozygous (rib1/+ or ribP7/+) embryos. (D) In situ hybridization analysis of SG genes repressed by Rib and with nearby Rib binding sites, in rib mutant and heterozygous embryos. (E) In situ hybridization analysis of a SG gene with nearby Rib binding sites whose expression is not detectably changed in rib mutant compared with heterozygous embryos. Rib mutants were identified by morphological criteria and/or the absence of expression of lacZ from the ftz-lacZ containing balancer chromosomes. Black arrowheads indicate SGs and white arrowheads indicate lacZ expression from the balancer chromosome in C-E.
... Results from DAVID clustering analysis of GO terms for genes repressed by Ribbon based on microarray data and bound by Ribbon in the salivary gland based on ChIP-Seq.
... Microarray gene expression analysis indicates the direction of transcriptional control of Rib targets. (A) RNA was isolated from three individual samples each of stage 11–16 WT and rib1/ribP7 embryos. Volcano plot shows genes that were downregulated (blue) or upregulated (red) at least 1.5-fold (P0.05) are indicated by gray. (B, C) Venn diagrams representing the overlap of 494 genes from the ChIP-seq and microarray (774 targets activated and 1176 repressed by Rib, respectively) data sets are shown. (D) The set of transcripts that are downregulated (blue) or upregulated (red) at least 1.5-fold (PChIP-seq analyses) are marked (cyan). (E, F) qRT-PCR results for a subset of genes obtained from the overlap of ChIP-seq and microarray data confirms significant expression change in the same direction as observed with microarray analysis for all but two examples, Sema-5C and CLS. *P<0.01, **P<0.001, Mann–Whitney U test.
Contributors:Xiao A. Huang, Hang Yin, Sarah Sweeney, Debasish Raha, Michael Snyder, Haifan Lin
Genome-wide Colocalization of Piwi and Piwi-Associated piRNAs
(A) Genome-wide localization of Piwi and Piwi-associated piRNAs are shown for X, 2L, 2R, 3L, 3R, 4, and unassembled Contig U. The horizontal line shows the length of each chromosome arm proportionally. The red peak above the chromosomal line represents Piwi ChIP scores [sum of ChIP-seq (U+M) scores per 10 kb window; both unique-mapping and multiple-mapping reads were considered], and the green peak below the chromosomal line represents the abundance of Piwi-associated piRNAs (numbers of piRNAs per 10 kb window). Gray ovals indicate centromeres. Asterisks denote enrichment of Piwi in telomere regions. Boxed regions labeled as B, C, and D are 260 kb region containing the 42AB piRNA cluster, a 150 kb region of sporadic transposons, and a 75 kb region containing a gene CG32377, respectively. See also Figure S2.
(B–D) Zoomed-in views of localization of Piwi and Piwi-associated piRNAs at the 42AB piRNA cluster (B), a sporadic transposon region (C), and CG32377 (D).
... Distinct Colocalization Patterns of Piwi and Piwi-Associated piRNAs in Euchromatin and Heterochromatin Suggest Two Modes of Piwi-piRNA Guidance Mechanism
(A and B) Relative positions of Piwi, Piwi-associated piRNA, and transposons within the euchromatic genome (X, 2L, 2R, 3L, 3R, 4) and the heterochromatic genome (XHet, 2LHet, 2RHet, 3LHet, 3RHet, YHet, U, and Uextra). Piwi-associated piRNAs were aligned at their 5′ ends in the same direction. Gray dash lines indicate positions of piRNAs. Piwi ChIP-seq scores within upstream and downstream regions surrounding piRNA-transcribing regions (±3 kb) were separately plotted for euchromatic genome (orange) and heterochromatic genome (blue), together with the transposon density (TE density; green). See also Figure S4.
(C and D) Relative positions of Piwi with piRNAs derived from piRNA clusters (green) and other sporadic piRNAs (orange).
(E) Heatmaps depict Piwi ChIP-seq scores over various types/classes of transposons within genome. Average Piwi ChIP-seq scores of all same types of transposons within genome (total) or only within piRNA clusters (cluster) were separately calculated.
(F) Heatmaps depict levels of chromatin-associated RNA Pol II over various types/classes of transposons within wild-type flies (left) and piwi1/piwi2 mutants (right). Average RNA Pol II ChIP-seq scores were separately calculated for all same types of transposons on chromosomal arms (X, 2L, 2R, 3L, 3R, 4), on contigs (U, Uextra) as well as for all same types of transposon within genome (genome average) or within piRNA clusters (cluster average).
... Distribution of ChIP-Seq (U+M) Scores over Genomic Features
Distribution of ChIP-seq (U+M) scores over CDS, 5′ UTR, 3′ UTR, introns, transposons, repetitive sequences, and intergenic regions within the whole genome (X, 2L, 2R, 3L, 3R, 4, XHet, 2LHet, 2RHet, 3LHet, 3RHet, YHet, U, Uextra). The top and bottom rows show the distributions in wild-type and piwi1/piwi2 flies, respectively. See also Figure S1 and Table S1.
... Chromosome-wide Changes of Chromatin States in Piwi Mutants
(A) Distribution of various epigenetic regulators/marks over the entire chromosome arm 2L [ChIP-seq (U+M) scores; both unique-mapping and repetitive sequences were considered] in wild-type and piwi1/piwi2 mutants. cen, the centromeric end of 2L; tel, the telomeric end of 2L. ChIP-seq scores were binned and averaged for every 20 kb window on the plots. See also Figures S5 and S6 and Tables S2 and S3.
(B) Distribution of various epigenetic regulators/marks over the entire contig Uextra (ChIP-seq [U+M] scores; both unique-mapping and repetitive sequences were considered) in wild-type and piwi1/piwi2 mutants. See also Figures S5 and S6 and Tables S2 and S3.
Contributors:Sarah E. Lacher, Joslynn S. Lee, Xuting Wang, Michelle R. Campbell, Douglas A. Bell, Matthew Slattery
Enriched GO categories for human orthologs of Drosophila class I genes.
... Direct Nrf2 regulation of enhancers at Keap1 and SQSTM1 in human cells. (A) Human Nrf2 ChIP-seq peak and ARE sequence at the Keap1 locus; gene models and tested enhancer regions are indicated as described for Fig. 4A. (B) Human Nrf2 ChIP-seq peak and ARE sequence at the SQSTM1 locus; gene models and tested enhancer regions are indicated as described for Fig. 4A. (C) EMSA as described for Fig. 4B, with AREs from Keap1 and SQSTM1 enhancers used as cold competitors; both are able to compete with labeled NQO1 probe in an ARE-dependent manner (lost with mutation of ARE), though the Keap1 ARE is a weaker competitor than the SQSTM1 ARE. (D) Luciferase reporter assay as described for Fig. 4C, but with enhancer from the Keap1 locus. (E) Same as (D) with enhancer from the SQSTM1 locus. Both the Keap1 (D) and the SQSTM1 (E) enhancers are upregulated in Nrf2+ in an ARE-dependent manner.
... Deeply conserved human Nrf2 targets are upregulated by sulforaphane in human cells. (A) Percentage of Nrf2 target genes overlapping human orthologs of Drosophila class I, II, or III genes. Orthologs were identified using either the top-scoring ortholog only (best ortholog) or all orthologs scoring >2 as described in the text. ⁎p Drosophila class I genes) and gene expression changes after treatment of LCL cells with sulforaphane (SFN).
... Enhancers at deeply conserved human target genes are regulated by Nrf2 in human cells. (A) Human Nrf2 ChIP-seq signal from LCL cells treated with DMSO or SFN as indicated. Select ancient Nrf2 target genes with highly significant binding are represented (ChIP y-axis scale, 0–5000). (B) ChIP-seq signal as in (A) at select ancient Nrf2 target genes with moderate binding (ChIP y-axis scale, 0–500). (C) Heat map representing the response to SFN, tert-butylhydroquinone (tBHQ), overexpression of Nrf2, or overexpression of a dominant negative version of Nrf2 (Nrf2DN) for reporter constructs driven by the enhancer regions highlighted in (A) and (B). NQO1 is a positive control for human Nrf2, but is not a conserved target because insects do not have an orthologous gene; the remaining nine are enhancers at deeply conserved Nrf2 target genes.
... Drosophila... Drosophila class III genes and human orthologs.
Contributors:Joshi, Anagha, Beck, Yvonne, Michoel, Tom
Transcription factors and their number of target genes in the D. melanogaster ChIP-chip gold standard network and in the predicted networks for six Drosophila species at the 10% recall level (in brackets for each TF the number of true positive predictions). The bottom two rows are the total number of interactions in each network and the overall precision (percentage of true positives) of the predicted networks.... We collected gene expression data for 3,610 genes in six Drosophila species measured at 9–13 time points during early embryonic development with 3–8 replicates per time point (200 samples in total) . To obtain a global view on the similarities and differences between samples, we performed multi-dimensional scaling using Sammon’s nonlinear mapping criterion on the 3,610-dimensional sample vectors (cf. Methods and Figure fig:sammona). The first (horizontal) axis of variation corresponded to developmental time, with samples ordered along this dimension according to increasing developmental time points, while the second (vertical) axis of variation corresponded to evolutionary distance, with samples ordered along this dimension according to species. By expanding these two axes of variation into principal components, we found that the “developmental” dimension explained 34% of the total variation in the data, while the “evolutionary” dimension explained 11% (cf. Methods). This result confirms that variations in gene expression levels across Drosophila species at the same developmental time point are not greater than variations across time points within the same species. In this study we were interested whether this additional layer of inter-species expression variation can be harnessed in the reconstruction of gene regulatory networks.... To quantitatively compare different methods across different gold standard networks we considered the area under the recall–precision curve (AUC) and the precision at 10% recall (PREC10) as performance measures and converted them to P -values by comparison to AUCs and PREC10s of networks generated by randomly assigning ranks to all possible edges in the corresponding gold standard network (cf. Methods and Figure fig:rec-prec-aggr for the recall vs. precision curves). While the AUC assesses the overall performance of a predicted network, PREC10 measures the quality of the top-ranked predictions, a property that may be of greater practical relevance. This analysis showed that no predicted network performs best for either measure across all gold standards (Figure fig:scorea-f). The single-species virilis networks performed best for 5 out of 12 AUC and PREC10 scores, albeit not for the ChIP-seq network measured in its own species. This overall good performance is consistent with virilis having the highest number of measured time points in the data (Supplementary Table tab:data). D. melanogaster also had more data points available than the other four species, but its time series were less complete (Supplementary Table tab:data). Among the integrative methods, the centroid and union methods both performed best for 5 out of 12 AUC and PREC10 scores (Figure fig:scorea-f). Both also had higher average AUC score than the best single-species network, but only the centroid method had higher average PREC10 score than the best single-species network (Figure fig:scoreg). The most important result however is the fact that the single-species network for the species were the gold standard network was measured never has the highest single-species AUC and only twice has the highest PREC10. In contrast, the centroid method always performs as good, and in most cases better, than the single-species network for the reference species (Figure fig:scorea-f). We conclude that the centroid method is the most robust network integration method achieving consistently high AUC and PREC10 scores, at least on this dataset.... Embryonic developmental time-course expression data in 6 Drosophila species (D. melanogaster (“amel”), D. ananassae (“ana”), D. persimilis (“per”), D. pseudoobscura (“pse”), D. simulans (“sim”) and D. virilis (“vir”)) was obtained from (ArrayExpress accession code E-MTAB-404). The data consists of 10 (amel), 13 (vir) or 9 (ana, per, pse, sim) developmental time points with several replicates per time point resulting in a total of 56 (amel), 36 (vir) or 27 (ana, per, pse, sim) arrays per species (Supplementary Table tab:data). The downloaded data was processed by averaging absolute expression levels over all reporters for a gene followed by taking the log 2 transform.... a. Number of interactions found in one to six species in the inferred gene regulatory networks at 10% recall level (red dots) and in 100 randomized networks with the same in- and out-degree distribution as the inferred networks (boxplots). b. Percentage of all predicted interactions (yellow) and of all true positive predictions (blue) in one to six species c. Precision of interactions found in one to six species. d. Recall of ChIP-seq gold standard interactions conserved in one to three species (green; data for BCD, KR, HB) and one to four species (red; data for TWI). e. Phylogenetic tree between six Drosophila species reconstructed from the inferred interactions at 10% recall level, with the total number of interactions in each species shown in brackets. The tree correctly splits the species in 3 groups – melanogaster (top), obscura (middle), virilis (bottom). Each branch, (numbered 1–9) represents a inferred network state transition. At each network state transition, the number of interactions inferred to be gained (red) or lost (blue) as well as the bootstrap value for each branch (in brackets) is indicated.... Although the gold standard network reconstructed from ChIP-chip data was in D. melanogaster, perhaps surprisingly the D. melanogaster predicted network did not perform better overall than the networks predicted for the other species (Figure fig:sammonb). To get confidence in this observation, we downloaded ChIP-sequencing data for three TFs (BCD, KR, HB) in three Drosophila species (melanogaster, pseudoobscura and virilis) and one TF (TWI) in four species (melanogaster, simulans, ananassae and pseudoobscura) , and created ChIP-seq gold standard networks for five species (cf. Methods). The recall-precision curves generated from the D. melanogaster ChIP-seq gold standard network (Supplementary Figure fig:rec-prec-singleb) were in good agreement with the ChIP-chip data, demonstrating again that the D. melanogaster predicted network performed no better than other Drosophila species. We also calculated recall-precision curves using the D. ananassae, D. pseudobscura, D. simulans and D. virilis ChIP-seq gold standard networks. Again, the regulatory network in that species did not perform better compared to the other species (Supplementary Figure fig:rec-prec-singlec–f).... Performance scores with respect to the gold standard ChIP-chip network for 14 TFs in D. melanogaster (a) and the ChIP-seq networks for D. melanogaster (b, 4 TFs), D. ananassae (c, 1 TF), D. pseudoobscura (d, 4 TFs), D. simulans (e, 1 TF), D. virilis (f, 4 TFs), and their averages over all gold standard networks (g). In each panel, the left, resp. right, figure shows - log 10 P A U C , resp. - log 10 P P R E C 10 for the six single-species predicted networks (green) and the five prediction aggregation methods (red). The dashed lines indicate the performance level of the single-species network for the gold standard species (a–f) or of the best performing single-species network (g). Values with a ∗ in panel a indicate numerical underflow values truncated to the smallest non-zero P -value ( 10 -324 ).... Recall vs. precision curves for predicted regulatory networks for five multi-species meta-analysis methods. The gold standard networks were the ChIP-chip network for 14 TFs in D. melanogaster (a) and the ChIP-seq networks for D. melanogaster (b, 4 TFs), D. ananassae (c, 1 TF), D. pseudoobscura (d, 4 TFs), D. simulans (e, 1 TF) and D. virilis (f, 4 TFs). The numbers in each legend are the area under the curve for each method.... Recall vs. precision curves for predicted regulatory networks in six Drosophila species. The gold standard networks were the ChIP-chip network for 14 TFs in D. melanogaster (a) and the ChIP-seq networks for D. melanogaster (b, 4 TFs), D. ananassae (c, 1 TF), D. pseudoobscura (d, 4 TFs), D. simulans (e, 1 TF) and D. virilis (f, 4 TFs). In panel a, the numbers in the legend are the area under the curve for each species. In panel b–f, the curve for the reference species is in red while the other species are in black.
Quantifying the distances between Giant, Hunchback and Kruppel ChIP-seq profiles and the profiles derived with the analytical model which includes DNA accessibility data. This is the same as Figure fig:heatmapChIPseq_GT_HB_KR_NoaccGroup070, except that we included binary DNA accessibility data in the analytical model. fig:heatmapChIPseq_GT_HB_KR_AccRegionsGroup070... Quantifying the distances between Bicoid and Caudal ChIP-seq profiles and the profiles derived with the analytical model. We plotted heatmaps for the correlation ( A ) and ( B ) and mean squared error ( C ) and ( D ) between the analytical model and the ChIP-seq profile of Bicoid ( A , C ) and Caudal ( B , D ). We computed these values for different sets of parameters: N ∈ 1 10 6 and λ ∈ 0.25 5 . We considered only the sites that have a PWM score higher than 70 % of the difference between the lowest and the highest score. ( A , B ) Orange colour indicates high correlation between the analytical model and the ChIP-seq profile, while white colour low correlation. ( C , D ) Blue colour indicates low mean squared error between the analytical model and the ChIP-seq profile, while white colour high mean squared error. ( E , F ) We plotted the regions where the mean square error is in the lower 12 % of the range of values (blue) and the correlation is the higher 12 % of the range of values (orange). With green rectangle we marked the optimal set of parameters in terms of mean squared error and with a black rectangle the intersection of the parameters for which the two regions intersect. fig:heatmapChIPseq_BCD_CAD_NoaccGroup070... First, there is an inconsistency in the experimental data in the sense there are peaks in the ChIP-seq profile that are located in DNA inaccessible areas, e.g. there are peaks in the Bicoid ChIP-seq profile at run, slp, eve, tll, gt, oc loci that overlap with DNA that is marked as inaccessible; see Figure fig:profileAllPositivesBCD in the Appendix. This indicates that either or both the DNA accessibility or the ChIP-seq data display some technical biases, e.g. , and, in these cases, the analytical model assumes that the DNA accessibility data is accurate and predicts that there is no binding in DNA inaccessible areas. One solution is to use continuous data for DNA accessibility, where different areas display different levels of accessibility. When using continuous values for DNA accessibility data, we did not observe any improvements of our model’s predictions. Nevertheless, we still observed ChIP-seq peaks for all five TFs that were overlapping with regions with reduce or no accessibility, thus, indicating the one or both data sets (ChIP-seq or DNase I) contain experimental biases; e.g. .... Nevertheless, Figures fig:profileAllPositivesHB and fig:profileAllPositivesKR in the Appendix show that the ChIP-seq profiles of Hunchback and Kruppel display some sharp peaks, which suggest that these two TFs display higher specificity than predicted by our approach. This contradicts our findings and one explanation for the few narrow ChIP-seq peaks is that these two TFs bind cooperatively to the genome. In this scenario, in the few narrow peaks for Hunchback and Kruppel, these TFs co-localise with co-factor(s) and previous studies identified that this is the case for both TFs; e.g. . This means that, by using our model, one could potentially underestimate the number of peaks in the binding profile.... The influence of weak binding on Hunchback and Kruppel ChIP-seq profiles. We plotted heatmaps for the correlation ( A ) and ( B ) and mean squared error ( C ) and ( D ) between the analytical model and the ChIP-seq profile of Hunchback A C and Kruppel B D . The analytical model includes binary DNA accessibility data (the accessibility of any site can be either 0 or 1 depending on whether the site is accessible or not). We computed these values for different sets of parameters: N ∈ 1 10 6 and λ ∈ 0.25 5 . Colour code as above. PWM filtering as in Figure fig:heatmapChIPseq_BCD_CAD_GT_AccRegionsGroup030. ( E , F ) We plotted the regions where the mean squared error is in the lower 12 % of the range of values (blue) and the correlation is the higher 12 % of the range of values (orange). With green rectangle we marked the optimal set of parameters in terms of mean squared error and with a black rectangle the intersection of the parameters for which the two regions intersect. fig:heatmapChIPseq_HB_KR_AccRegionsGroup030... Genome-wide quality of the fit. The boxplots represent the A C correlation and B D mean squared error between the ChIP-seq data sets and the analytically estimated profiles. We partitioned the genome in 20 K b p regions and we kept only the regions that had at least one DNA accessible site ( 4599 regions). Next for each ChIP-seq data set we selected the regions where the mean ChIP-seq signal is higher than a proportion of the background (see Table tab:ChIPseqProfileStatistics in the Appendix). In A B , we selected the regions with a mean ChIP-seq signal higher than the background ( > B ). In C D , we selected the regions with a mean ChIP-seq signal higher than half the background ( > 0.5 ⋅ B ). The numbers of DNA regions that display a mean ChIP-seq signal higher than the thresholds are listed in Table ... Quantifying the distances between Bicoid and Caudal ChIP-seq profiles and the profiles derived with the analytical model which includes DNA accessibility data. This is the same as Figure fig:heatmapChIPseq_BCD_CAD_NoaccGroup070, except that we included binary DNA accessibility data in the analytical model. fig:heatmapChIPseq_BCD_CAD_AccRegionsGroup070... Binding profiles for Hunchback at all 21 loci. The grey shading represents a ChIP-seq profile, the red line represents the prediction of the analytical model, the yellow shading represents the inaccessible DNA and the vertical blue lines represent the percentage of occupancy of the site (we only displayed sites with an occupancy higher than 5 % ). We considered the optimal set of parameters for Hunchback ( 2000 m o l e c u l e s and λ = 3.00 ).... One advantage of our analytical model is that it can be used to predict the binding profiles genome-wide and, thus, we extended the analysis from the original twenty one loci to the entire genome. We partitioned the genome in 20 K b p regions, from which we removed regions that did not have any accessible site. For each ChIP-seq profile, we then selected the regions that display a ChIP-seq signal higher than the genome-wide background. We found that the quality of our model’s predictions vary widely; see Figure fig:genomeWideQuality ( A ) and ( B ). In particular, there are regions where the correlation between our model predictions and the ChIP-seq profile is high, but at the same time regions where this correlation is low.... Kaplan et al. found that, at loci with low binding (low ChIP-seq signal), the correlation between the statistical thermodynamics model and the ChIP-seq profile was low. To test whether this is valid genome-wide, we also analysed regions where the mean signal is higher than half of the genome-wide background (leading in an increase in the number of investigated loci). Our results confirm that there is a decrease in the mean correlation when including regions with lower ChIP-seq signal; see Figure fig:genomeWideQuality ( C ). We also perform a Kolmogorov-Smirnov test that showed that in the case of Bicoid and Caudal this difference is statistically significant; see Figure fig:GenomeWideKSPvalue in the Appendix. This also means that, at least for regions with strong binding, the model predictions are highly correlated with the ChIP-seq profile as previously found ; see Figure fig:genomeWideQuality. Nevertheless, for regions with low binding, in addition to the reduction in the correlation we also observed a decrease in the mean squared error, which is statistically significant in the case of Bicoid, Caudal and Kruppel; see Figure fig:GenomeWideKSPvalue in the Appendix. Note that for Giant and Hunchback the difference is not statistically significant due to the small number of loci included in the analysis; see Table tab:GenomeWideNoOfRegions in the Appendix. This indicates that our model is able to correctly capture the low signal in those regions, but there is little or no correlation to the actual ChIP-seq signal. One explanation for this result is that, in those regions, there is little or no binding and what the ChIP-seq method recovers might be considered technical noise.
ChIP-seq... The mouse and Drosophila PcG targets are associated with different sequence signatures. (A) The Pho and Yy1 motifs are similar. (B) ROC curves and corresponding AUC scores for cross-specifies prediction.
... The Hox clusters are targeted by PcG in both mouse and Drosophila but their promoter sequences show different properties. (A) The Drosophila ANT-C region. (B) The mouse Hoxb cluster. Each colored box represents a protein-coding gene, in the order of their chromosomal locations. The TSS coordinates of the genes are shown as vertical lines in the bottom of the figure. The color indicates either presence (red) or absence (blue) of a specific feature labeled on the left. The locations of the Hox genes are marked in the above. The label “:ChIP” after certain TFs is used to indicate that target information is based on ChIP-chip data.
... Enrichment analysis for overlaps between TF and PcG targets. (A) The 15 TFs probed by ChIP-chip/seq experiments in mouse ESCs. The statistically significant one are marked by asterisks (p<1.0E-7 from one-sided Fisher exact test with Bonferroni correction). (B) The most enriched or depleted TF motifs.
... Predicted propensity scores reflect the overall PcG target plasticity. (A) Comparison of the propensity score distribution among different gene groups with similar H3K27me3 profiles. The number of lineages in which the genes are marked by H3K27me3 is shown above the figure. The number of genes in each group is also shown (in parentheses). (B) Time course gene expression level analysis. The PcG target genes in ESCs are divided into 15 roughly equal-sized groups associated with similar propensity scores (mean values shown on the left). The heat map indicates the mean mRNA expression level within each group at different time points after LIF removal. (C) Comparison of the propensity score distributions for the Ezh2-/-, H3K27me3+ and Ezh2-/-, H3K27me3- genes, which correspond to the subset of PcG targets that either retain or lose the H3K27me3 mark in the Ezh2-/- mutant ESCs. (D) Enrichment score for overlap between the top 18 TF features and Ezh2-/-, H3K27me3+ or Ezh2-/-, H3K27me3- targets. The label “:ChIP” after certain TFs is used to indicate that target information is based on ChIP-chip/seq data. The enrichment score is defined as the ratio of the observed frequency of a TF feature among PcG targets over the frequency expected by chance.
... Predicted propensity scores and PcG status in Drosophila.
Contributors:Tamer Ali, Rainer Renkawitz, Marek Bartkuhn
Insulators, chromatin domains and topologically associated domains (TADs). Interaction matrix representing a virtual Hi-C experiment (top). The grey scale above indicates interaction frequencies. Interactions occur predominantly within TADs (e.g. enhancer–promoter interactions), which are often grouped in subdomains. Interactions between TAD boundaries are thought to depend on the binding of CTCF (shown as a schematic ChIP-seq track in red) to its cognate DNA-binding motif (black arrows). CTCF sites not involved in binding to TAD boundaries are shown in pale red and grey motifs, respectively. Motifs involved in long-range chromatin interactions show an inverted repeat orientation (see Figure 2). As not all TAD boundaries are bound by CTCF it is likely that additional factors may be involved in their function (indicated by question mark). TADs are often co-incident with chromatin domains represented by a schematic ChIP-seq track for an active (H3K36me3; green) and a repressive (H3K27me3; blue) histone modification. Active TADs are gene-rich (black bars for active genes) in contrast to gene-poor repressed domains (grey bars).
... Insulator components with conserved features in vertebrates and Drosophila.
... Drosophila CP190 recruitment and strength of TAD boundaries/insulators correlate with combinatorial binding of architectural proteins. (a) The interaction matrix represents TADs. Boundaries between TADs are often marked by CP190 binding (schematic ChIP-seq track, blue). CP190 is recruited to chromatin by a wide variety of insulator binding factors (IBPs, as exemplified by CTCF, BEAF32 and Pita in schematic ChIP-seq tracks). Frequently, different insulator binding factors cluster together, suggesting a cooperative recruitment mode for targeting CP190 to chromatin. Combinatorial recruitment of CP190 to TAD boundaries may be functionally important since high occupancy of IBPs and other architectural proteins such as cohesin, condensin and TFIIIC predict the strength of insulator function as well as TAD borders [54••]. It should be noted that not all TAD boundaries are bound by known IBPs (?) and that many IBP binding sites are found within TADs. (b) The physical DNA string model summarizes the contact and binding data illustrated in (a).