Contents

1 Introduction

ChIP-seq has been widely utilized as the standard technology to detect protein binding regions, where peak calling algorithms were developed particularly to serve the analysis. Existing peak callers lack of power on ranking peaks’ significance due to sequencing technology might undergo sequence context biases, e.g. GC bias. gcapc is designed to address this deficiency by modeling GC effects into peak calling. gcapc can also help refine the significance of peaks called by other peak callers, or correct the GC-content bias for a read count table for a predefined set of genomic regions across a series of samples. The gcapc package requires the inputs as one ChIP-seq BAM file (for peak calling/refining) or a read count table (for GC effects removal) as well as other optional parameters.

A common analysis for peak calling/refining contains four steps.

  1. Reads coverage. In this step, BAM file records will be converted to coverages on basepair resolution for forward and reverse strands separately.

  2. Binding width estimation. This parameter is a measurement for the size of protein binding region in crosslinked complexes of ChIP experiments. Also, peak detection half size are estimated based on region signals from two strands.

  3. GC effects estimation. Generalized linear mixture models followed by EM algorithms are performed to evaluate potential GC effects.

  4. Peak calling/refining. For peak calling, enrichment scores are evaluated by permutation analysis for significance. Peaks are reported with enrichment scores and p-values. For peak refining, peaks called by other peak callers should be provided as a GRanges object. New enrichment significances are added as meta columns for the input peaks.

For correcting GC effects on a count table, one step analysis based on function refineSites is enough.

2 Getting Started

Load the package in R

library(gcapc)

3 Preparing Inputs

Preparing inputs for correcting GC effects on a count table should be easy by referring to the function man page. Here, we focus on inputs for peak calling and refining. The inputs could be as minimum as a path to a BAM file, which is an indexed alignment records for sequencing reads. However, additional options are encouraged to be specified to accelerate the analysis and improve the accuracy. The following set are the options which can be customized by users.

  1. BAM records filtering options. In the function read5endCoverage, reads can be filtered for selected chromosomes, mapping quality, duplicate removal, etc. Downstream analysis could be highly accelerated if only a subset of chromosomes are analyzed. This actually suggests a divide and conquer strategy if one ChIP-seq experiment is extremely deeply sequenced. In that case, analysis based on each chromosome level could save lots of memory.

  2. Sequencing fragments options. If one has prior knowledge on the size of sequencing fragments. The optional arguments in function bindWidth could be specified to limit searching in narrower ranges; Or, this function can be omitted if binding width are known in advance. Note that this binding width might not be equivalent to the binding width of protein in biology, since it could be affected by crosslinking operations.

  3. Sampling size for GC effects estimation. The default is 0.05, which means 5% of genome will be used if analysis is based on whole genome. However, for smaller genomes or small subset of chromosomes, this size should be tuned higher to ensure accuracy. Or, sample genome multiple times, and use average estimation to aviod sampling bias. Note: larger sample size or more sampling times result longer computation of GC effects estimation.

  4. GC ranges for GC effects estimation. As illustrated in the man pages, GC ranges (gcrange parameter) should be carefully selected. The reason is that regions with extremely low/high GC content sometimes act as outliers, and can drive the regression lines when selected forground regions are too few in mixture model fitting. This happens when the studied binding protein has too few genome-wide binding events.

  5. EM algorithm priors and convergence. Options for EM algorithms can be tuned to accelerate the iterations.

  6. Permutation times. As we suggested in the function help page, a proper times of permutation could save time as well as ensuring accuracy.

In this vignette, we will use enbedded file chipseq.bam as one example to illustrate this package. This file contains about ~80000 reads from human chromosome 21 for CTCF ChIP-seq data.

bam <- system.file("extdata", "chipseq.bam", package="gcapc")

4 Peak Calling/Refining

For details of the algorithms, please refer to our paper (Teng and Irizarry 2017).

4.1 Reads coverage

The first step is to generate the reads coverage for both forward and reverse strands. The coverage is based on single nucleotide resolution and uses only the 5’ ends of BAM records. That means, if duplicates are not allowed, the maximum coverage for every nucleotide is 1.

cov <- read5endCoverage(bam)
cov
## $fwd
## RleList of length 1
## $chr21
## integer-Rle of length 48129895 with 40225 runs
##   Lengths: 9414767       1    8350       1 ...     116       1   41437
##   Values :       0       1       0       1 ...       0       1       0
## 
## 
## $rev
## RleList of length 1
## $chr21
## integer-Rle of length 48129895 with 40427 runs
##   Lengths: 9412972       1    3087       1 ...     367       1   34767
##   Values :       0       1       0       1 ...       0       1       0

Obejct cov is a two-element list representing coverages for forward and reverse strands, respectively, while each element is a list for coverages on individual chromosomes.

4.2 Binding width

The second step is to estimate the binding width and peak detection half window size of ChIP-seq experiment. This step could be omitted if binding width is known in advance. Binding width is further treated as the size of region unit for effective GC bias estimation and peak calling. Peak detection half window size is used to define width of flanking regions.

If additional information is known from sequencing fragments, this step could be speeded up. For example, narrowing down the range size helps.

bdw <- bindWidth(cov, range=c(50L,300L), step=10L)
## Starting to estimate bdwidth.
## ...... Cycle 1 for bind width estimation
## ...... Cycle 2 for bind width estimation
## ...... Cycle 3 for bind width estimation
## ...... Cycle final for bind width estimation
## ...... Estimated bind width as 111 
## ...... Estimated peak window half size as 220 
## Refining peak window half size by region from two strands
## ......................
## ...... Refined peak window half size as 155
bdw
## [1] 111 155

4.3 GC effects

This step performs GC effects estimation using the proposed models. It is noted that by allowing to display the plots, one can view intermediate results which provide you direct sense on your ChIP-seq data, such as the extent of GC effects. Also, the EM algorithms iterations are enabled by default to display the trace of log likelihood changes, and other notification messages are printed for courtesy.

layout(matrix(1:2,1,2))
gcb <- gcEffects(cov, bdw, sampling=c(0.25,1), plot=TRUE, model='poisson')
## Starting to estimate GC effects
## ...... Too few/much ranges in 'supervise'. Selecting random sampling
## ......... Sampling 25 percent of regions by 1 times, total 108395 regions
## ...... Counting reads
## ...... Calculating GC content with flanking 99 
## ...... Estimating GC effects
## ......... Iteration 1    ll -39354.64    increment 107573.1 
## ......... Iteration 2    ll -38490.04    increment 864.6023 
## ......... Iteration 3    ll -38435.08    increment 54.95739 
## ......... Iteration 4    ll -38439.09    increment -4.012721

Here, 25% of windows were sampled with 1 repeat. The left figure provides the correlation between forward and reverse strands signals, by using the estimated binding width as region unit. The right figure shows the raw and predicted GC effects using mixture model.

Two other options need to be noted here are supervise and model. If supervise option is specified as a GRanges object, it provides a set of potential peaks and allows more efficient sampling procedure. In detail, the two mixtures are sampled separately from forground (signal) and background regions. model option allows switching between Poisson and Negative Binomial distribution (default) in model fitting. Theoratically, Negative Binomial assumption is more accurate than poisson. Nevertheless, Poisson is a good approximation to Negative Binomial for GC effect estimation here, and shows much faster computing speed than Negative Binomial especially when the total number of selected bins is large.

4.4 Peak significance

This is the last step for peak calling. It uses information generated in previous steps, calculates enrichment scores and performs permutation analysis to propose significant peak regions. Final peaks are formated into GRanges object, and meta columns are used to record significance. Additional notification messages are also printed.

layout(matrix(1:2,1,2))
peaks <- gcapcPeaks(cov, gcb, bdw, plot=TRUE, permute=100L)
peaks <- gcapcPeaks(cov, gcb, bdw, plot=TRUE, permute=50L)

peaks
## GRanges object with 431 ranges and 2 metadata columns:
##         seqnames               ranges strand |        es
##            <Rle>            <IRanges>  <Rle> | <numeric>
##     [1]    chr21 [ 9825469,  9825566]      * |     7.929
##     [2]    chr21 [ 9827175,  9827489]      * |    24.968
##     [3]    chr21 [ 9909588,  9909776]      * |     8.344
##     [4]    chr21 [11068738, 11068840]      * |     5.608
##     [5]    chr21 [11175032, 11175180]      * |     6.594
##     ...      ...                  ...    ... .       ...
##   [427]    chr21 [47705666, 47705851]      * |     7.281
##   [428]    chr21 [47745288, 47745462]      * |    12.284
##   [429]    chr21 [48055027, 48055217]      * |    14.052
##   [430]    chr21 [48059795, 48059983]      * |     10.84
##   [431]    chr21 [48081159, 48081322]      * |    32.315
##                           pv
##                    <numeric>
##     [1]   0.0156410425983946
##     [2] 8.26671736919986e-08
##     [3]   0.0126251787677631
##     [4]   0.0484566038671704
##     [5]   0.0305940463101507
##     ...                  ...
##   [427]   0.0217632081476766
##   [428]  0.00135706432332783
##   [429] 0.000423751932345229
##   [430]  0.00317342746368843
##   [431] 8.26671736919986e-08
##   -------
##   seqinfo: 1 sequence from an unspecified genome; no seqlengths

It is noted that here two tests using different number of permutation times results almost the same cutoff on enrichment scores, which suggests small number of permutations are allowed to save time. The left figure shows here the cutoff on enrichment scores based on 50 times of permutations, and right figure shows it based on 100 times of permutations. Note that we only used chromosome 21 for illustration, thus increased permutation times from default 5 to 50 here.

4.5 Peak refining

In order to remove GC effects on other peak callers’ outputs, this package provides function to refine enrichment significance for given peaks. Peaks have to be provided as a GRanges object. A flexible set of peaks are preferred to reduce potential false negative, meaning both significant (e.g. p<=0.05) and non-significant (e.g. p>0.05) peaks are preferred to be included. If the total number of peaks is not too big, a reasonable set of peaks include all those with p-value/FDR less than 0.99 by other peak callers.

newpeaks <- refinePeaks(cov, gcb, bdw, peaks=peaks, permute=50L)
## Warning: In BioC 3.5, the 'force' argument was replaced by the more
##   flexible 'pruning.mode' argument, and is deprecated. See ?seqinfo
##   for the supported pruning modes. Note that 'force=TRUE' is
##   equivalent to 'pruning.mode="coarse"'.
plot(newpeaks$es,newpeaks$newes,xlab='old score',ylab='new score')

newpeaks
## GRanges object with 431 ranges and 4 metadata columns:
##         seqnames               ranges strand |        es
##            <Rle>            <IRanges>  <Rle> | <numeric>
##     [1]    chr21 [ 9825469,  9825566]      * |     7.929
##     [2]    chr21 [ 9827175,  9827489]      * |    24.968
##     [3]    chr21 [ 9909588,  9909776]      * |     8.344
##     [4]    chr21 [11068738, 11068840]      * |     5.608
##     [5]    chr21 [11175032, 11175180]      * |     6.594
##     ...      ...                  ...    ... .       ...
##   [427]    chr21 [47705666, 47705851]      * |     7.281
##   [428]    chr21 [47745288, 47745462]      * |    12.284
##   [429]    chr21 [48055027, 48055217]      * |    14.052
##   [430]    chr21 [48059795, 48059983]      * |     10.84
##   [431]    chr21 [48081159, 48081322]      * |    32.315
##                           pv     newes                newpv
##                    <numeric> <numeric>            <numeric>
##     [1]   0.0156410425983946     7.929   0.0328913572872633
##     [2] 8.26671736919986e-08    24.968  1.2635544937023e-06
##     [3]   0.0126251787677631     8.344   0.0275125206763462
##     [4]   0.0484566038671704     5.608   0.0821378193346811
##     [5]   0.0305940463101507     6.594   0.0567554217974637
##     ...                  ...       ...                  ...
##   [427]   0.0217632081476766     7.281   0.0431948401029223
##   [428]  0.00135706432332783    12.284  0.00417283128101453
##   [429] 0.000423751932345229    14.052  0.00157082797279917
##   [430]  0.00317342746368843     10.84  0.00879629204190402
##   [431] 8.26671736919986e-08    32.315 1.14868590332659e-07
##   -------
##   seqinfo: 1 sequence from an unspecified genome; no seqlengths

Here, two new meta columns are added into previous peaks (GRanges), including adjusted significance and p-values. It is noted that the new enrichment scores are actually the same as previously calculated (figure above) since the peak regions were previously called by gcapc. In practice, peak regions and their significances called by other peak callers mostly would be different from gcapc if there is strong GC bias. If refining those peak significances, the improvement should be obvious as we showed in our paper.

5 Correcting GC Effects for A Count Table

In this package, function refineSites is provided in case that some are more interested in correcting GC effects for pre-defined regions instead of peak calling. We used this function to adjust signals for ENCODE reported sites in our paper (Teng and Irizarry 2017). This function is eay-to-use, with a count table and corresponding genomic regions are the two required inputs. For detail of using this function, please read the function man page.

6 Summary

In this vignette, we went through main functions in this package, and illustrated how they worked. By following these steps, users will be able to remove potential GC effects either for peak identification or for read count signals.

References

Teng, Mingxiang, and Rafael A. Irizarry. 2017. “Accounting for Gc-Content Bias Reduces Systematic Errors and Batch Effects in Chip-Seq Data.” bioRxiv.