Common Downstream ChIP-seq Analysis Workflows using ChIPpeakAnno
26 July 2019
Package
ChIPpeakAnno 3.18.2
Contents
- 1 Workflow for ChIP-seq experiments of single transcription factor with replicates
- 1.1 Import data and obtain overlapping peaks from replicates
- 1.2 Prepare annotation data
- 1.3 Visualize binding site distribution relative to features
- 1.4 Annotate peaks
- 1.5 Obtain enriched GO terms and Pathways
- 1.6 Obtain the sequences surrounding the peaks
- 1.7 Output a summary of consensus in the peaks
- 1.8 Find peaks with bi-directional promoters
- 1.9 Find possible enhancers with DNA interaction data
- 2 Workflow for comparing binding profiles from multiple transcription factors (TFs)
In this guide, we illustrate here two common downstream analysis workflows for ChIP-seq experiments, one is for comparing and combining peaks for single transcription factor (TF) with replicates, and the other is for comparing binding profiles from ChIP-seq experiments with multiple TFs.
1 Workflow for ChIP-seq experiments of single transcription factor with replicates
This workflow shows how to convert BED/GFF files to GRanges, find overlapping peaks between two peak sets, and visualize the number of common and specific peaks with Venn diagram.
1.1 Import data and obtain overlapping peaks from replicates
The input for ChIPpeakAnno1 is a list of called peaks identified
from ChIP-seq experiments or any other experiments that yield a set of chromosome coordinates. Although peaks are represented as GRanges
in ChIPpeakAnno, other common peak formats such as BED, GFF and MACS can be converted to GRanges easily using a conversion toGRanges method. For detailed information on how to use this method, please type ?toGRanges.
The following examples illustrate the usage of this method to convert BED and GFF file to GRanges, add metadata from orignal peaks to the overlap GRanges using function addMetadata, and visualize the overlapping using function makeVennDiagram.
library(ChIPpeakAnno)
bed <- system.file("extdata", "MACS_output.bed", package="ChIPpeakAnno")
gr1 <- toGRanges(bed, format="BED", header=FALSE)
## one can also try import from rtracklayer
gff <- system.file("extdata", "GFF_peaks.gff", package="ChIPpeakAnno")
gr2 <- toGRanges(gff, format="GFF", header=FALSE, skip=3)
## must keep the class exactly same as gr1$score, i.e., numeric.
gr2$score <- as.numeric(gr2$score)
ol <- findOverlapsOfPeaks(gr1, gr2)
## add metadata (mean of score) to the overlapping peaks
ol <- addMetadata(ol, colNames="score", FUN=mean)
ol$peaklist[["gr1///gr2"]][1:2]## GRanges object with 2 ranges and 2 metadata columns:
## seqnames ranges strand | peakNames
## <Rle> <IRanges> <Rle> | <CharacterList>
## [1] chr1 713791-715578 * | gr1__MACS_peak_13,gr2__001,gr2__002
## [2] chr1 724851-727191 * | gr2__003,gr1__MACS_peak_14
## .mcols[, colNames]
## <numeric>
## [1] 850.203333333333
## [2] 29.17
## -------
## seqinfo: 1 sequence from an unspecified genome; no seqlengthsmakeVennDiagram(ol)
Figure 1: Figure 1
Venn diagram of overlaps for replicated experiments
## $p.value
## gr1 gr2 pval
## [1,] 1 1 0
##
## $vennCounts
## gr1 gr2 Counts
## [1,] 0 0 0
## [2,] 0 1 61
## [3,] 1 0 62
## [4,] 1 1 166
## attr(,"class")
## [1] "VennCounts"1.2 Prepare annotation data
Annotation data should be an object of GRanges. Same as import peaks, we use the method toGRanges, which can return an object of GRanges, to represent the annotation data. An annotation data be constructed from not only BED, GFF or user defined readable text files, but also EnsDb or TxDb object, by calling the toGRanges method. Please type ?toGRanges for more information.
library(EnsDb.Hsapiens.v75) ##(hg19)
## create annotation file from EnsDb or TxDb
annoData <- toGRanges(EnsDb.Hsapiens.v75, feature="gene")
annoData[1:2]## GRanges object with 2 ranges and 1 metadata column:
## seqnames ranges strand | gene_name
## <Rle> <IRanges> <Rle> | <character>
## ENSG00000223972 chr1 11869-14412 + | DDX11L1
## ENSG00000227232 chr1 14363-29806 - | WASH7P
## -------
## seqinfo: 273 sequences from GRCh37 genome1.3 Visualize binding site distribution relative to features
After finding the overlapping peaks, the distribution of the distance of overlapped peaks to the nearest feature such as the transcription start sites (TSS) can be plotted by binOverFeature function.
The sample code here plots the distribution of peaks around the TSS.
overlaps <- ol$peaklist[["gr1///gr2"]]
binOverFeature(overlaps, annotationData=annoData,
radius=5000, nbins=20, FUN=length, errFun=0,
ylab="count",
main="Distribution of aggregated peak numbers around TSS")
Figure 2: Figure 2
Distribution of peaks around transcript start sites.
In addition, assignChromosomeRegion can be used to summarize the distribution of peaks over different type of features such as exon, intron, enhancer, proximal promoter, 5’ UTR and 3’ UTR.
This distribution can be summarized in peak centric or nucleotide centric view using
the function assignChromosomeRegion.
Please note that one peak might span multiple type of features, leading to the number of annotated features greater than the total number of input peaks. At the peak centric view, precedence will dictate the annotation order when peaks span multiple type of features.
library(TxDb.Hsapiens.UCSC.hg19.knownGene)
aCR<-assignChromosomeRegion(overlaps, nucleotideLevel=FALSE,
precedence=c("Promoters", "immediateDownstream",
"fiveUTRs", "threeUTRs",
"Exons", "Introns"),
TxDb=TxDb.Hsapiens.UCSC.hg19.knownGene)
barplot(aCR$percentage, las=3)
Figure 3: Figure 3
Peak distribution over different genomic features.
1.4 Annotate peaks
As shown from the distribution of aggregated peak numbers around TSS and
the distribution of peaks in different of chromosome regions,
most of the peaks locate around TSS. Therefore, it is reasonable to use annotatePeakInBatch or annoPeaks
to annotate the peaks to the promoter regions of Hg19 genes.
Promoters can be specified with bindingRegion. For the following example, promoter region is defined as upstream 2000 and downstream 500 from TSS (bindingRegion=c(-2000, 500)).
overlaps.anno <- annotatePeakInBatch(overlaps,
AnnotationData=annoData,
output="nearestBiDirectionalPromoters",
bindingRegion=c(-2000, 500))
library(org.Hs.eg.db)
overlaps.anno <- addGeneIDs(overlaps.anno,
"org.Hs.eg.db",
IDs2Add = "entrez_id")
head(overlaps.anno)## GRanges object with 6 ranges and 11 metadata columns:
## seqnames ranges strand | peakNames
## <Rle> <IRanges> <Rle> | <CharacterList>
## X1 chr1 713791-715578 * | gr1__MACS_peak_13,gr2__001,gr2__002
## X1 chr1 713791-715578 * | gr1__MACS_peak_13,gr2__001,gr2__002
## X3 chr1 839467-840090 * | gr1__MACS_peak_16,gr2__004
## X4 chr1 856361-856999 * | gr1__MACS_peak_17,gr2__005
## X5 chr1 859315-860144 * | gr2__006,gr1__MACS_peak_18
## X10 chr1 901118-902861 * | gr2__012,gr1__MACS_peak_23
## .mcols[, colNames] peak feature feature.ranges
## <numeric> <character> <character> <IRanges>
## X1 850.203333333333 X1 ENSG00000228327 700237-714006
## X1 850.203333333333 X1 ENSG00000237491 714150-745440
## X3 73.12 X3 ENSG00000272438 840214-851356
## X4 54.69 X4 ENSG00000223764 852245-856396
## X5 81.485 X5 ENSG00000187634 860260-879955
## X10 119.91 X10 ENSG00000187583 901877-911245
## feature.strand distance insideFeature distanceToSite gene_name
## <Rle> <integer> <factor> <integer> <character>
## X1 - 0 overlapStart 0 RP11-206L10.2
## X1 + 0 overlapStart 0 RP11-206L10.9
## X3 + 123 upstream 123 RP11-54O7.16
## X4 - 0 overlapStart 0 RP11-54O7.3
## X5 + 115 upstream 115 SAMD11
## X10 + 0 overlapStart 0 PLEKHN1
## entrez_id
## <character>
## X1 <NA>
## X1 <NA>
## X3 <NA>
## X4 100130417
## X5 148398
## X10 84069
## -------
## seqinfo: 1 sequence from an unspecified genome; no seqlengthswrite.csv(as.data.frame(unname(overlaps.anno)), "anno.csv")The distribution of the common peaks around features can be visualized using a pie chart.
pie1(table(overlaps.anno$insideFeature))
Figure 4: Figure 4
Pie chart of the distribution of common peaks around features.
1.5 Obtain enriched GO terms and Pathways
The following example shows how to use getEnrichedGO to obtain a list of enriched GO terms with annotated peaks. For pathway analysis, please use function getEnrichedPATH with reactome or KEGG database. Please note that by default feature_id_type is set as “ensembl_gene_id”. If you are using TxDb as annotation data, please set it to “entrez_id”.
over <- getEnrichedGO(overlaps.anno, orgAnn="org.Hs.eg.db",
maxP=.05, minGOterm=10,
multiAdjMethod="BH", condense=TRUE)
head(over[["bp"]][, -c(3, 10)])## [1] go.id go.term Ontology
## [4] pvalue count.InDataset count.InGenome
## [7] totaltermInDataset totaltermInGenome EntrezID
## <0 rows> (or 0-length row.names)library(reactome.db)
path <- getEnrichedPATH(overlaps.anno, "org.Hs.eg.db", "reactome.db", maxP=.05)
head(path)## path.id EntrezID count.InDataset count.InGenome pvalue
## 1 R-HSA-1296041 2782 1 25 0.04323569
## 2 R-HSA-1296059 2782 1 25 0.04323569
## 3 R-HSA-196854 375790 2 189 0.04445965
## 4 R-HSA-196854 55229 2 189 0.04445965
## 5 R-HSA-1971475 375790 1 26 0.04492589
## 6 R-HSA-199220 55229 1 17 0.02960656
## totaltermInDataset totaltermInGenome PATH
## 1 194 109842 NA
## 2 194 109842 NA
## 3 194 109842 NA
## 4 194 109842 NA
## 5 194 109842 NA
## 6 194 109842 NA1.6 Obtain the sequences surrounding the peaks
Here is an example to get the sequences of the peaks plus 20 bp upstream and downstream for PCR validation or motif discovery.
library(BSgenome.Hsapiens.UCSC.hg19)
seq <- getAllPeakSequence(overlaps, upstream=20, downstream=20, genome=Hsapiens)
write2FASTA(seq, "test.fa")1.7 Output a summary of consensus in the peaks
Here is an example to get the Z-scores for short oligos3.
## summary of the short oligos
freqs <- oligoFrequency(Hsapiens$chr1, MarkovOrder=3)
os <- oligoSummary(seq, oligoLength=6, MarkovOrder=3,
quickMotif=FALSE, freqs=freqs)
## plot the results
zscore <- sort(os$zscore)
h <- hist(zscore, breaks=100, xlim=c(-50, 50), main="Histogram of Z-score")
text(zscore[length(zscore)], max(h$counts)/10,
labels=names(zscore[length(zscore)]), adj=1)
Figure 5: Figure 5
Histogram of Z-score of 6-mer
## We can also try simulation data
seq.sim.motif <- list(c("t", "g", "c", "a", "t", "g"),
c("g", "c", "a", "t", "g", "c"))
set.seed(1)
seq.sim <- sapply(sample(c(2, 1, 0), 1000, replace=TRUE, prob=c(0.07, 0.1, 0.83)),
function(x){
s <- sample(c("a", "c", "g", "t"),
sample(100:1000, 1), replace=TRUE)
if(x>0){
si <- sample.int(length(s), 1)
if(si>length(s)-6) si <- length(s)-6
s[si:(si+5)] <- seq.sim.motif[[x]]
}
paste(s, collapse="")
})
os <- oligoSummary(seq.sim, oligoLength=6, MarkovOrder=3,
quickMotif=TRUE)
zscore <- sort(os$zscore, decreasing=TRUE)
h <- hist(zscore, breaks=100, main="Histogram of Z-score")
text(zscore[1:2], rep(5, 2),
labels=names(zscore[1:2]), adj=0, srt=90)
Figure 6: Figure 6
Histogram of Z-score of simulation data
## generate the motifs
library(motifStack)
pfms <- mapply(function(.ele, id)
new("pfm", mat=.ele, name=paste("SAMPLE motif", id)),
os$motifs, 1:length(os$motifs))
motifStack(pfms[[1]])
(#fig:simulation.motif)Figure 7. Motif of simulation data
1.8 Find peaks with bi-directional promoters
Bidirectional promoters are the DNA regions located between TSS of two adjacent genes that are transcribed on opposite directions and often co-regulated by this shared promoter region5. Here is an example to find peaks near bi-directional promoters.
bdp <- peaksNearBDP(overlaps, annoData, maxgap=5000)
c(bdp$percentPeaksWithBDP,
bdp$n.peaks,
bdp$n.peaksWithBDP)## [1] 0.1084337 166.0000000 18.0000000bdp$peaksWithBDP[1:2]## GRangesList object of length 2:
## $1
## GRanges object with 2 ranges and 11 metadata columns:
## seqnames ranges strand | peakNames
## <Rle> <IRanges> <Rle> | <CharacterList>
## X1 chr1 713791-715578 * | gr1__MACS_peak_13,gr2__001,gr2__002
## X1 chr1 713791-715578 * | gr1__MACS_peak_13,gr2__001,gr2__002
## .mcols[, colNames] bdp_idx peak feature feature.ranges
## <numeric> <integer> <character> <character> <IRanges>
## X1 850.203333333333 1 X1 ENSG00000228327 700237-714006
## X1 850.203333333333 1 X1 ENSG00000237491 714150-745440
## feature.strand distance insideFeature distanceToSite gene_name
## <Rle> <integer> <factor> <integer> <character>
## X1 - 0 overlapStart 0 RP11-206L10.2
## X1 + 0 overlapStart 0 RP11-206L10.9
##
## $4
## GRanges object with 2 ranges and 11 metadata columns:
## seqnames ranges strand | peakNames
## X4 chr1 856361-856999 * | gr1__MACS_peak_17,gr2__005
## X4 chr1 856361-856999 * | gr1__MACS_peak_17,gr2__005
## .mcols[, colNames] bdp_idx peak feature feature.ranges
## X4 54.69 4 X4 ENSG00000223764 852245-856396
## X4 54.69 4 X4 ENSG00000187634 860260-879955
## feature.strand distance insideFeature distanceToSite gene_name
## X4 - 0 overlapStart 0 RP11-54O7.3
## X4 + 3260 upstream 3260 SAMD11
##
## -------
## seqinfo: 1 sequence from an unspecified genome; no seqlengths1.9 Find possible enhancers with DNA interaction data
There are several techniques available to determine the spatial organization of chromosomes at high resolution such as 3C, 5C and HiC6. These techniques make it possible to search peaks binding to the potential enhancer regions. Here is an example to find peaks binding to the potential enhancer regions.
DNA5C <- system.file("extdata",
"wgEncodeUmassDekker5CGm12878PkV2.bed.gz",
package="ChIPpeakAnno")
DNAinteractiveData <- toGRanges(gzfile(DNA5C))
findEnhancers(overlaps, annoData, DNAinteractiveData)## GRanges object with 5 ranges and 14 metadata columns:
## seqnames ranges strand | peakNames
## <Rle> <IRanges> <Rle> | <CharacterList>
## X1 chr1 151591700-151591800 * | gr2__229,gr1__MACS_peak_229
## X1 chr1 151591700-151591800 * | gr2__229,gr1__MACS_peak_229
## X1 chr1 151591700-151591800 * | gr2__229,gr1__MACS_peak_229
## X1 chr1 151591700-151591800 * | gr2__229,gr1__MACS_peak_229
## X1 chr1 151630186-151630286 * | gr2__230,gr1__MACS_peak_230
## .mcols[, colNames] feature feature.ranges feature.strand
## <numeric> <character> <IRanges> <Rle>
## X1 78.675 ENSG00000207606 151518272-151518367 +
## X1 78.675 ENSG00000143390 151313116-151319833 -
## X1 78.675 ENSG00000143376 151584541-151671567 +
## X1 78.675 ENSG00000143367 151512781-151556059 +
## X1 78.675 ENSG00000143393 151264273-151300191 -
## feature.shift.ranges feature.shift.strand distance insideFeature
## <IRanges> <Rle> <integer> <factor>
## X1 151594534-151594629 + 2733 upstream
## X1 151595209-151601927 + 3408 upstream
## X1 151500588-151587615 - 4084 upstream
## X1 151595902-151639180 + 4101 upstream
## X1 151594247-151630165 - 20 upstream
## distanceToSite gene_name peak DNAinteractive.ranges
## <integer> <character> <character> <IRanges>
## X1 2733 MIR554 X1 151516086-151603110
## X1 3408 RFX5 X1 151309062-151603110
## X1 4084 SNX27 X1 151546428-151636526
## X1 4101 TUFT1 X1 151546428-151636526
## X1 20 PI4KB X1 151283079-151636526
## DNAinteractive.blocks
## <IRangesList>
## X1 1-19082,76263-87025
## X1 1-13633,283287-294049
## X1 1-6978,72324-90099
## X1 1-6978,72324-90099
## X1 1-5699,335673-353448
## -------
## seqinfo: 1 sequence from an unspecified genome; no seqlengths2 Workflow for comparing binding profiles from multiple transcription factors (TFs)
Given two or more peak lists from different TFs, one may be interested in finding whether DNA binding profile of those TFs are correlated, and if correlated, what is the common binding pattern. The workflow here shows how to test the correlation of binding profiles of three TFs and how to discover the common binding pattern.
2.1 Import data
path <- system.file("extdata", package="ChIPpeakAnno")
files <- dir(path, "broadPeak")
data <- sapply(file.path(path, files), toGRanges, format="broadPeak")
names(data) <- gsub(".broadPeak", "", files)2.2 Determine if there is a significant overlap among multiple sets of peaks
2.2.1 Hypergeometric test
When we test the association between two sets of data based on hypergeometric
distribution, the number of all potential binding sites is required.
The parameter totalTest in the function makeVennDiagram indicates how many
potential peaks in total will be used in the hypergeometric test. It should be
larger than the largest number of peaks in the peak list. The smaller it is
set, the more stringent the test is. The time used to calculate p-value does not
depend on the value of the totalTest.
For practical guidance on how to choose totalTest,
please refer to the post.
The following example makes an assumption that there are 3% of coding region plus promoter region. Because the sample data is only a subset of chromosome 2,
we estimate that the total binding sites is 1/24 of possible binding region in the genome.
ol <- findOverlapsOfPeaks(data, connectedPeaks="keepAll")
averagePeakWidth <- mean(width(unlist(GRangesList(ol$peaklist))))
tot <- ceiling(3.3e+9 * .03 / averagePeakWidth / 24)
makeVennDiagram(ol, totalTest=tot, connectedPeaks="keepAll")
Figure 7: Figure 8
Venn diagram of overlaps.
## $p.value
## TAF Tead4 YY1 pval
## [1,] 0 1 1 1.000000e+00
## [2,] 1 0 1 2.904297e-258
## [3,] 1 1 0 8.970986e-04
##
## $vennCounts
## TAF Tead4 YY1 Counts count.TAF count.Tead4 count.YY1
## [1,] 0 0 0 849 0 0 0
## [2,] 0 0 1 621 0 0 621
## [3,] 0 1 0 2097 0 2097 0
## [4,] 0 1 1 309 0 310 319
## [5,] 1 0 0 59 59 0 0
## [6,] 1 0 1 166 172 0 170
## [7,] 1 1 0 8 8 8 0
## [8,] 1 1 1 476 545 537 521
## attr(,"class")
## [1] "VennCounts"2.2.2 Permutation test
The above hypergeometric test requires users to input an estimate of
the total potential binding sites for a given TF. To circumvent this requirement, we implemented a permutation test called peakPermTest.
Before performing a permutation test, users need to generate random peaks using the distribution discovered from the input peaks for a given feature type (transcripts or exons), to make sure the binding positions relative to features, such as TSS and geneEnd, and the width of the random peaks follow the distribution of that of the input peaks.
Alternatively, a peak pool representing all potential
binding sites can be created with associated binding probabilities for random peak sampling using preparePool. Here is an example to build a peak pool for human genome using the transcription factor binding site clusters (V3) (see ?wgEncodeTfbsV3)
downloaded from ENCODE with the HOT spots (?HOT.spots) removed. HOT spots are the genomic regions with high probability of being bound by many TFs in ChIP-seq experiments7.
We suggest remove those HOT spots from the peak lists before performing
permutation test to avoid the overestimation of the association between the two input
peak lists. Users can also choose to remove ENCODE blacklist for a given species.
The blacklists were constructed by identifying consistently problematic regions over independent cell lines and types of experiments for each species in the ENCODE and modENCODE datasets8.
Please note that some of the blacklists may need to be converted to the correct genome assembly using liftover utility.
Following are the sample codes to do the permutation test using permTest:
data(HOT.spots)
data(wgEncodeTfbsV3)
hotGR <- reduce(unlist(HOT.spots))
removeOl <- function(.ele){
ol <- findOverlaps(.ele, hotGR)
if(length(ol)>0) .ele <- .ele[-unique(queryHits(ol))]
.ele
}
TAF <- removeOl(data[["TAF"]])
TEAD4 <- removeOl(data[["Tead4"]])
YY1 <- removeOl(data[["YY1"]])
# we subset the pool to save demo time
set.seed(1)
wgEncodeTfbsV3.subset <-
wgEncodeTfbsV3[sample.int(length(wgEncodeTfbsV3), 2000)]
pool <- new("permPool", grs=GRangesList(wgEncodeTfbsV3.subset), N=length(YY1))
pt1 <- peakPermTest(YY1, TEAD4, pool=pool, seed=1, force.parallel=FALSE)
plot(pt1)
Figure 8: Figure 9
permutation test for YY1 and TEAD4
pt2 <- peakPermTest(YY1, TAF, pool=pool, seed=1, force.parallel=FALSE)
plot(pt2)
Figure 9: Figure 10
permutation test for YY1 and TAF
2.3 Visualize and compare the binding pattern
The binding pattern around a genome feature could be visualized and compared using a side-by-side heatmap and density plot using the binding ranges of overlapping peaks.
features <- ol$peaklist[[length(ol$peaklist)]]
feature.recentered <- reCenterPeaks(features, width=4000)
## here we also suggest importData function in bioconductor trackViewer package
## to import the coverage.
## compare rtracklayer, it will save you time when handle huge dataset.
library(rtracklayer)
files <- dir(path, "bigWig")
if(.Platform$OS.type != "windows"){
cvglists <- sapply(file.path(path, files), import,
format="BigWig",
which=feature.recentered,
as="RleList")
}else{## rtracklayer can not import bigWig files on Windows
load(file.path(path, "cvglist.rds"))
}
names(cvglists) <- gsub(".bigWig", "", files)
feature.center <- reCenterPeaks(features, width=1)
sig <- featureAlignedSignal(cvglists, feature.center,
upstream=2000, downstream=2000)
##Because the bw file is only a subset of the original file,
##the signals are not exists for every peak.
keep <- rowSums(sig[[2]]) > 0
sig <- sapply(sig, function(.ele) .ele[keep, ], simplify = FALSE)
feature.center <- feature.center[keep]
heatmap <- featureAlignedHeatmap(sig, feature.center,
upstream=2000, downstream=2000,
upper.extreme=c(3,.5,4))
Figure 10: Figure 11
Heatmap of aligned features sorted by signal of TAF
sig.rowsums <- sapply(sig, rowSums, na.rm=TRUE)
d <- dist(sig.rowsums)
hc <- hclust(d)
feature.center$order <- hc$order
heatmap <- featureAlignedHeatmap(sig, feature.center,
upstream=2000, downstream=2000,
upper.extreme=c(3,.5,4),
sortBy="order")
Figure 11: Figure 12
Heatmap of aligned features sorted by hclut
featureAlignedDistribution(sig, feature.center,
upstream=2000, downstream=2000,
type="l")
Figure 12: Figure 13
Distribution of aligned features
1. Zhu, L. J. et al. ChIPpeakAnno: A bioconductor package to annotate chip-seq and chip-chip data. BMC bioinformatics 11, 237 (2010).
2. Zhu, L. J. in Tiling arrays 105–124 (Springer, 2013).
3. Leung, M.-Y., Marsh, G. M. & Speed, T. P. Over-and underrepresentation of short dna words in herpesvirus genomes. Journal of Computational Biology 3, 345–360 (1996).
4. Helden, J. van, Olmo, M. lí del & Pérez-Ortín, J. E. Statistical analysis of yeast genomic downstream sequences reveals putative polyadenylation signals. Nucleic Acids Research 28, 1000–1010 (2000).
5. Robertson, G. et al. Genome-wide profiles of stat1 dna association using chromatin immunoprecipitation and massively parallel sequencing. Nature methods 4, 651–657 (2007).
6. Lieberman-Aiden, E. et al. Comprehensive mapping of long-range interactions reveals folding principles of the human genome. science 326, 289–293 (2009).
7. Yip, K. Y. et al. Classification of human genomic regions based on experimentally determined binding sites of more than 100 transcription-related factors. Genome Biol 13, R48 (2012).
8. Consortium, E. P. & others. An integrated encyclopedia of dna elements in the human genome. Nature 489, 57–74 (2012).