ChIP-seq

Kharchenko et al. (2008).

• Tags versus sequenced reads; single-end read extension in 3’ direction
• Strand shift / cross-correlation
• Defined (narrow, e.g., transcription factor binding sites) versus diffuse (e.g., histone marks) peaks

ChIP-seq for differential binding

• Designed experiment with replicate samples per treatment
• Analysis using insights from microarrays / RNA-seq

Novel statistical issues

• Inferring peaks without ‘data snooping’ (using the same data twice, once to infer peaks, once to estimate differential binding)
• Retaining power
• Minimizing false discovery rate

Work flow

• Following Bailey et al., 2013

Experimental design and execution

• Single sample

  • ChIPed transcription factor and
  • Input (fragmented genomic DNA) or control (e.g., IP with non-specific antibody such as immunoglobulin G, IgG)

• Designed experiments

  • Replication of TF / control pairs

Sequencing & alignment

• Sequencing depth rules of thumb: \(>10M\) reads for narrow peaks, \(>20M\) for broad peaks
• Long & paired end useful but not essential – alignment in ambiguous regions
• Basic aligners generally adequate, e.g., no need to align splice junctions
• Sims et al., 2014

Peak calling

• Very large number of peak calling programs; some specialized for e.g., narrow vs. broad peaks.
• Commmonly used: MACS, PeakSeq, CisGenome, …

• MACS: Model-based Analysis for ChIP-Seq, Liu et al., 2008

  • Scale control tag counts to match ChIP counts
  • Center peaks by shifting \(d/2\)
  • Model occurrence of a tag as a Poisson process
  • Look for fixed width sliding windows with exceess number of tag enrichment
  • Empirical FDR: Swap ChIP and control samples; FDR is # control peaks / # ChIP peaks
  • Output: BED file of called peaks

Down-stream analysis

• Annotation: what genes are my peaks near?
• Differential representation: which peaks are over- or under-represented in treatment 1, compared to treatment 2?
• Motif identification (peaks over known motifs?) and discovery
• Integrative analysis, e.g., assoication of regulatory elements and expression

Peak calling

‘Known’ ranges

• Count tags in pre-defined ranges, e.g., promoter regions of known genes
• Obvious limitations, e.g., regulatory elements not in specified ranges; specified range contains multiple regulatory elements with complementary behavior

de novo windows

• Width: narrow peaks, 1bp; broad peaks, 150bp
• Offset: 25-100bp; influencing computational burden

de novo peak calling

• Third-party software (many available; MACS commonly used)

• Various strategies for calling peaks – Lun & Smyth, Table 1

  • Call each sample independently; intersection or union of peaks across samples, …
  • Call peaks from a pooled library
  • …

• Relevant slides pdf

Peak calling across libraries

• Table 1: Description of peak calling strategies. Each strategy is given an identifier and is described by the mode in which MACS is run, the libraries on which it is run and the consolidation operation (if any) performed to combine peaks between libraries or groups. For method 6, the union of the peaks in each direction of enrichment is taken.

ID	Mode	Library	Operation
1	Single-sample	Individual	Union
2	Single-sample	Individual	Intersection
3	Single-sample	Individual	At least 2
4	Single-sample	Pooled over group	Union
5	Single-sample	Pooled over group	Intersection
6	Two-sample	Pooled over group	Union
7	Single-sample	Pooled over all	–

• How to choose? – Lun & Smyth,

  • Under the null hypothesis, type I error rate is uniform
  • Table 2: consequences for type I error
  • Best strategy: call peaks from a pooled library
  • Table 2: The observed type I error rate when testing for differential enrichment using counts from each peak calling strategy. Error rates for a range of specified error thresholds are shown. All values represent the mean of 10 simulation iterations with the standard error shown in brackets. RA: reference analysis using 10 000 randomly chosen true peaks.

ID	Error rate
	0.01	0.05	0.1
RA	0.010 (0.000)	0.051 (0.001)	0.100 (0.002)
1	0.002 (0.000)	0.019 (0.001)	0.053 (0.001)
2	0.003 (0.000)	0.030 (0.000)	0.073 (0.001)
3	0.006 (0.000)	0.042 (0.001)	0.092 (0.001)
4	0.033 (0.001)	0.145 (0.001)	0.261 (0.002)
5	0.000 (0.000)	0.001 (0.000)	0.005 (0.000)
6	0.088 (0.006)	0.528 (0.013)	0.893 (0.006)
7	0.010 (0.000)	0.049 (0.001)	0.098 (0.001)

## 100,000 t-tests under the null, n = 6
n <- 6; m <- matrix(rnorm(n * 100000), ncol=n)
P <- genefilter::rowttests(m, factor(rep(1:2, each=3)))$p.value
quantile(P, c(.001, .01, .05))

##    0.1%      1%      5% 
## 0.00109 0.01013 0.05035

hist(P, breaks=20)

de novo hybrid strategies

1 - 4: Experimental Design … Alignment

The experiment involves changes in binding profiles of the NFYA protein between embryonic stem cells and terminal neurons. It is a subset of the data provided by Tiwari et al. 2012 available as GSE25532. There are two es (embryonic stem cell) and two tn (terminal neuron) replicates. Single-end FASTQ files were extracted from GEO, aligned using Rsubread, and post-processed (sorted and indexed) using Rsamtools with the script available at

system.file(package="UseBioconductor", "scripts", "ChIPSeq", "NFYA", 
    "preprocess.R")

Create a data frame summarizing the files used.

files <- local({
    acc <- c(es_1="SRR074398", es_2="SRR074399", tn_1="SRR074417",
             tn_2="SRR074418")
    data.frame(Treatment=sub("_.*", "", names(acc)),
               Replicate=sub(".*_", "", names(acc)),
               sra=sprintf("%s.sra", acc),
               fastq=sprintf("%s.fastq.gz", acc),
               bam=sprintf("%s.fastq.gz.subread.BAM", acc),
               row.names=acc, stringsAsFactors=FALSE)
})

5: Reduction

Change to the directory where the BAM files are located

setwd("~/UseBioconductor-data/ChIPSeq/NFYA")

Load the csaw library and count reads in overlapping windows. This returns a SummarizedExperiment, so explore it a bit…

library(csaw)
library(GenomicRanges)
frag.len <- 110
system.time({
    data <- windowCounts(files$bam, width=10, ext=frag.len)
})                                      # 156 seconds
acc <- sub(".fastq.*", "", data$bam.files)
colData(data) <- cbind(files[acc,], colData(data))

6: Analysis

Filtering (vignette Chapter 3) Start by filtering low-count windows. There are likely to be many of these (how many?). Is there a rational way to choose the filtering threshold?

library(edgeR)     # for aveLogCPM()
keep <- aveLogCPM(assay(data)) >= -1
data <- data[keep,]

Normalization (composition bias) (vignette Chapter 4) csaw uses binned counts in normalization. The bins are large relative to the ChIP peaks, on the assumption that the bins primarily represent non-differentially bound regions. The sample bin counts are normalized using the edgeR TMM (trimmed median of M values) method seen in the RNASeq differential expression lab. Explore vignette chapter 4 for more on normalization (this is a useful resource when seeking to develop normalization methods for other protocols!).

system.time({
    binned <- windowCounts(files$bam, bin=TRUE, width=10000)
})                                      #139 second
normfacs <- normalize(binned)

Experimental design and Differential binding (vignette Chapter 5) Differential binding will be assessed using edgeR, where we need to specify the experimental design

design <- model.matrix(~Treatment, colData(data))

Apply a standard edgeR work flow to identify differentially bound regions. Creatively explore the results.

y <- asDGEList(data, norm.factors=normfacs)
y <- estimateDisp(y, design)
fit <- glmQLFit(y, design, robust=TRUE)
results <- glmQLFTest(fit, contrast=c(0, 1))
head(results$table)

##    logFC logCPM     F PValue
## 1 -0.674  -1.36 0.631 0.4321
## 2 -0.769  -1.40 0.800 0.3772
## 3 -0.362  -1.33 0.186 0.6686
## 4  0.494  -1.37 0.346 0.5599
## 5  1.476  -1.17 3.135 0.0852
## 6  2.192  -1.24 5.763 0.0217

Multiple testing (vignette Chapter 6) The challenge is that FDR across all detected differentially bound regions is what one is interested in, but what is immediately available is the FDR across differentially bound windows; region will often consist of multiple overlapping windows. As a first step, we’ll take a ‘quick and dirty’ approach to identifying regions by merging ‘high-abundance’ windows that are within, e.g., 1kb of one another

merged <- mergeWindows(rowRanges(data), tol=1000L)

Combine test results across windows within regions. Several strategies are explored in section 6.5 of the vignette.

tabcom <- combineTests(merged$id, results$table)
head(tabcom)

##   nWindows logFC.up logFC.down PValue   FDR
## 1        1        0          1 0.4321 0.610
## 2        2        0          1 0.6686 0.804
## 3        4        3          0 0.0826 0.213
## 4        1        1          0 0.3898 0.571
## 5        2        0          2 0.1904 0.361
## 6        1        0          0 0.8320 0.913

Section 6.6 of the vignette discusses approaches to identifying the ‘best’ windows within regions.

Finally, create a GRangesList that associated with two result tables and the genomic ranges over which the results were calculated.

gr <- rowRanges(data)
mcols(gr) <- as(results$table, "DataFrame")
grl <- split(gr, merged$id)
mcols(grl) <- as(tabcom, "DataFrame")

Annotation