Package version: derfinder 1.10.6

Contents

1 Asking for help

Please read the basics of using derfinder in the quick start guide. Thank you.

2 derfinder users guide

If you haven’t already, please read the quick start to using derfinder vignette. It explains the basics of using derfinder, how to ask for help, and showcases an example analysis.

The derfinder users guide goes into more depth about what you can do with derfinder. It covers the two implementations of the DER Finder approach (Jaffe, Murakami, Lee, Leek, et al., 2012a). That is, the (A) expressed regions-level and (B) single base-level F-statistics implementations. The vignette also includes advanced material for fine tuning some options, working with non-human data, and example scripts for a high-performance computing cluster.

3 Expressed regions-level

The expressed regions-level implementation is based on summarizing the coverage information for all your samples and applying a cutoff to that summary. For example, by calculating the mean coverage at every base and then checking if it’s greater than some cutoff value of interest. Contiguous bases passing the cutoff become a candidate Expressed Region (ER). We can then construct a matrix summarizing the base-level coverage for each sample for the set of ERs. This matrix can then be using with packages developed for feature counting (limma, DESeq2, edgeR, etc) to determine which ERs have differential expression signal. That is, to identify the Differentially Expressed Regions (DERs).

3.1 regionMatrix()

Commonly, users have aligned their raw RNA-seq data and saved the alignments in BAM files. Some might have chosen to compress this information into BigWig files. BigWig files are much faster to load than BAM files and are the type of file we prefer to work with. However, we can still identify the expressed regions from BAM files.

The function regionMatrix() will require you to load the data (either from BAM or BigWig files) and store it all in memory. It will then calculate the mean coverage across all samples, apply the cutoff you chose, and determine the expressed regions.

This is the path you will want to follow in most scenarios.

3.2 railMatrix()

Rail logo

Our favorite method for identifying expressed regions is based on pre-computed summary coverage files (mean or median) as well as coverage files by sample. Rail is a cloud-enabled aligner that will generate:

  1. Per chromosome, a summary (mean or median) coverage BigWig file. Note that the summary is adjusted for library size and is by default set to libraries with 40 million reads.
  2. Per sample, a coverage BigWig file with the un-adjusted coverage.

Rail does a great job in creating these files for us, saving time and reducing the memory load needed for this type of analysis with R.

If you have this type of data or can generate it from BAM files with other tools, you will be interested in using the railMatrix() function. The output is identical to the one from regionMatrix(). It’s just much faster and memory efficient. The only drawback is that BigWig files are not fully supported in Windows as of rtracklayer version 1.25.16.

We highly recommend this approach. Rail has also other significant features such as: scalability, reduced redundancy, integrative analysis, mode agnosticism, and inexpensive cloud implementation. For more information, visit rail.bio.

4 Single base-level F-statistics

The DER Finder approach was originally implemented by calculating t-statistics between two groups and using a hidden markov model to determine the expression states: not expressed, expressed, differentially expressed (Jaffe, Murakami, Lee, Leek, et al., 2012a). The original software works but had areas where we worked to improve it, which lead to the single base-level F-statistics implementation. Also note that the original software is no longer maintained.

This type of analysis first loads the data and preprocess it in a format that saves time and reduces memory later. It then fits two nested models (an alternative and a null model) with the coverage information for every single base-pair of the genome. Using the two fitted models, it calculates an F-statistic. Basically, it generates a vector along the genome with F-statistics.

A cutoff is then applied to the F-statistics and contiguous base-pairs of the genome passing the cutoff are considered a candidate Differentially Expressed Region (DER).

Calculating F-statistics along the genome is computationally intensive, but doable. The major resource drain comes from assigning p-values to the DERs. To do so, we permute the model matrices and re-calculate the F-statistics generating a set of null DERs. Once we have enough null DERs, we compare the observed DERs against the null DERs to calculate p-values for the observed DERs.

This type of analysis at the chromosome level is done by the analyzeChr() function, which is a high level function using many other pieces of derfinder. Once all chromosomes have been processed, mergeResults() combines them.

Which implementation of the DER Finder approach you will want to use depends on your specific use case and computational resources available. But in general, we recommend starting with the expressed regions-level implementation.

5 Example data

In this vignette we we will analyze a small subset of the samples from the BrainSpan Atlas of the Human Brain (Jaffe, Murakami, Lee, Leek, et al., 2012b) publicly available data.

We first load the required packages.

## Load libraries
library('derfinder')
library('derfinderData')
library('GenomicRanges')

5.1 Phenotype data

For this example, we created a small table with the relevant phenotype data for 12 samples: 6 from fetal samples and 6 from adult samples. We chose at random a brain region, in this case the amygdaloid complex. For this example we will only look at data from chromosome 21. Other variables include the age in years and the gender. The data is shown below.

library('knitr')
## Get pheno table
pheno <- subset(brainspanPheno, structure_acronym == 'AMY')

## Display the main information
p <- pheno[, -which(colnames(pheno) %in% c('structure_acronym',
    'structure_name', 'file'))]
rownames(p) <- NULL
kable(p, format = 'html', row.names = TRUE)
gender lab Age group
1 F HSB97.AMY -0.4423077 fetal
2 M HSB92.AMY -0.3653846 fetal
3 M HSB178.AMY -0.4615385 fetal
4 M HSB159.AMY -0.3076923 fetal
5 F HSB153.AMY -0.5384615 fetal
6 F HSB113.AMY -0.5384615 fetal
7 F HSB130.AMY 21.0000000 adult
8 M HSB136.AMY 23.0000000 adult
9 F HSB126.AMY 30.0000000 adult
10 M HSB145.AMY 36.0000000 adult
11 M HSB123.AMY 37.0000000 adult
12 F HSB135.AMY 40.0000000 adult

5.2 Load the data

derfinder offers three functions related to loading raw data. The first one, rawFiles(), is a helper function for identifying the full paths to the input files. Next, loadCoverage() loads the base-level coverage data from either BAM or BigWig files for a specific chromosome. Finally, fullCoverage() will load the coverage for a set of chromosomes using loadCoverage().

We can load the data from derfinderData (Yin, Lawrence, and Cook, 2017) by first identifying the paths to the BigWig files with rawFiles() and then loading the data with fullCoverage(). Note that the BrainSpan data is already normalized by the total number of mapped reads in each sample. However, that won’t be the case with most data sets in which case you might want to use the totalMapped and targetSize arguments. The function getTotalMapped() will be helpful to get this information.

## Determine the files to use and fix the names
files <- rawFiles(system.file('extdata', 'AMY', package = 'derfinderData'),
    samplepatt = 'bw', fileterm = NULL)
names(files) <- gsub('.bw', '', names(files))

## Load the data from disk
system.time(fullCov <- fullCoverage(files = files, chrs = 'chr21',
    totalMapped = rep(1, length(files)), targetSize = 1))
## 2017-08-21 19:56:43 fullCoverage: processing chromosome chr21
## 2017-08-21 19:56:43 loadCoverage: finding chromosome lengths
## 2017-08-21 19:56:43 loadCoverage: loading BigWig file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB113.bw
## 2017-08-21 19:56:43 loadCoverage: loading BigWig file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB123.bw
## 2017-08-21 19:56:44 loadCoverage: loading BigWig file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB126.bw
## 2017-08-21 19:56:44 loadCoverage: loading BigWig file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB130.bw
## 2017-08-21 19:56:44 loadCoverage: loading BigWig file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB135.bw
## 2017-08-21 19:56:44 loadCoverage: loading BigWig file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB136.bw
## 2017-08-21 19:56:44 loadCoverage: loading BigWig file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB145.bw
## 2017-08-21 19:56:44 loadCoverage: loading BigWig file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB153.bw
## 2017-08-21 19:56:44 loadCoverage: loading BigWig file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB159.bw
## 2017-08-21 19:56:45 loadCoverage: loading BigWig file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB178.bw
## 2017-08-21 19:56:45 loadCoverage: loading BigWig file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB92.bw
## 2017-08-21 19:56:45 loadCoverage: loading BigWig file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB97.bw
## 2017-08-21 19:56:45 loadCoverage: applying the cutoff to the merged data
## 2017-08-21 19:56:45 filterData: normalizing coverage
## 2017-08-21 19:56:45 filterData: done normalizing coverage
## 2017-08-21 19:56:45 filterData: originally there were 48129895 rows, now there are 48129895 rows. Meaning that 0 percent was filtered.
##    user  system elapsed 
##   2.140   0.040   2.182

Alternatively, since the BigWig files are publicly available from BrainSpan (see here), we can extract the relevant coverage data using fullCoverage(). Note that as of rtracklayer 1.25.16 BigWig files are not supported on Windows: you can find the fullCov object inside derfinderData to follow the examples.

## Determine the files to use and fix the names
files <- pheno$file
names(files) <- gsub('.AMY', '', pheno$lab)

## Load the data from the web
system.time(fullCov <- fullCoverage(files = files, chrs = 'chr21',
    totalMapped = rep(1, length(files)), targetSize = 1))

Note how loading the coverage for 12 samples from the web was quite fast. Although in this case we only retained the information for chromosome 21.

The result of fullCov is a list with one element per chromosome. If no filtering was performed, each chromosome has a DataFrame with the number of rows equaling the number of bases in the chromosome with one column per sample.

## Lets explore it
fullCov
## $chr21
## DataFrame with 48129895 rows and 12 columns
##          HSB113 HSB123 HSB126 HSB130 HSB135 HSB136 HSB145 HSB153 HSB159
##           <Rle>  <Rle>  <Rle>  <Rle>  <Rle>  <Rle>  <Rle>  <Rle>  <Rle>
## 1             0      0      0      0      0      0      0      0      0
## 2             0      0      0      0      0      0      0      0      0
## 3             0      0      0      0      0      0      0      0      0
## 4             0      0      0      0      0      0      0      0      0
## 5             0      0      0      0      0      0      0      0      0
## ...         ...    ...    ...    ...    ...    ...    ...    ...    ...
## 48129891      0      0      0      0      0      0      0      0      0
## 48129892      0      0      0      0      0      0      0      0      0
## 48129893      0      0      0      0      0      0      0      0      0
## 48129894      0      0      0      0      0      0      0      0      0
## 48129895      0      0      0      0      0      0      0      0      0
##          HSB178 HSB92 HSB97
##           <Rle> <Rle> <Rle>
## 1             0     0     0
## 2             0     0     0
## 3             0     0     0
## 4             0     0     0
## 5             0     0     0
## ...         ...   ...   ...
## 48129891      0     0     0
## 48129892      0     0     0
## 48129893      0     0     0
## 48129894      0     0     0
## 48129895      0     0     0

If filtering was performed, each chromosome also has a logical Rle indicating which bases of the chromosome passed the filtering. This information is useful later on to map back the results to the genome coordinates.

5.3 Filter coverage

Depending on the use case, you might want to filter the base-level coverage at the time of reading it, or you might want to keep an unfiltered version. By default both loadCoverage() and fullCoverage() will not filter.

If you decide to filter, set the cutoff argument to a positive value. This will run filterData(). Note that you might want to standardize the library sizes prior to filtering if you didn’t already do it when creating the fullCov object. You can do so by by supplying the totalMapped and targetSize arguments to filterData(). Note: don’t use these arguments twice in fullCoverage() and filterData().

In this example, we prefer to keep both an unfiltered and filtered version. For the filtered version, we will retain the bases where at least one sample has coverage greater than 2.

## Filter coverage
filteredCov <- lapply(fullCov, filterData, cutoff = 2)
## 2017-08-21 19:56:46 filterData: originally there were 48129895 rows, now there are 130356 rows. Meaning that 99.73 percent was filtered.

The result is similar to fullCov but with the genomic position index as shown below.

## Similar to fullCov but with $position
filteredCov
## $chr21
## $chr21$coverage
## DataFrame with 130356 rows and 12 columns
##        HSB113 HSB123 HSB126 HSB130 HSB135 HSB136 HSB145 HSB153 HSB159
##         <Rle>  <Rle>  <Rle>  <Rle>  <Rle>  <Rle>  <Rle>  <Rle>  <Rle>
## 1        2.01      0   0.00   0.04   0.23   0.13      0   0.59   0.15
## 2        2.17      0   0.00   0.04   0.23   0.13      0   0.63   0.15
## 3        2.21      0   0.00   0.04   0.23   0.13      0   0.67   0.15
## 4        2.37      0   0.00   0.04   0.23   0.13      0   0.71   0.15
## 5        2.37      0   0.06   0.04   0.23   0.13      0   0.75   0.25
## ...       ...    ...    ...    ...    ...    ...    ...    ...    ...
## 130352   1.25   1.28   2.05   0.79   1.63   1.37   1.03   2.21   2.46
## 130353   1.21   1.24   2.05   0.79   1.63   1.37   1.03   2.21   2.46
## 130354   1.21   1.20   1.93   0.79   1.63   1.28   0.73   2.01   2.21
## 130355   1.17   1.20   1.93   0.79   1.63   1.28   0.60   2.01   2.11
## 130356   1.17   1.09   1.87   0.79   1.55   1.02   0.56   1.93   1.96
##        HSB178 HSB92 HSB97
##         <Rle> <Rle> <Rle>
## 1        0.05  0.07  1.58
## 2        0.05  0.07  1.58
## 3        0.05  0.10  1.61
## 4        0.05  0.13  1.65
## 5        0.10  0.13  1.68
## ...       ...   ...   ...
## 130352   2.07  2.23  1.51
## 130353   2.07  2.23  1.51
## 130354   2.11  2.13  1.47
## 130355   2.11  2.13  1.47
## 130356   1.78  2.06  1.47
## 
## $chr21$position
## logical-Rle of length 48129895 with 3235 runs
##   Lengths: 9825448     149       2       9 ...     740     598   45091
##   Values :   FALSE    TRUE   FALSE    TRUE ...   FALSE    TRUE   FALSE

In terms of memory, the filtered version requires less resources. Although this will depend on how rich the data set is and how aggressive was the filtering step.

## Compare the size in Mb
round(c(fullCov = object.size(fullCov),
    filteredCov = object.size(filteredCov)) / 1024^2, 1)
##     fullCov filteredCov 
##        22.7         8.5

Note that with your own data, filtering for bases where at least one sample has coverage greater than 2 might not make sense: maybe you need a higher or lower filter. The amount of bases remaining after filtering will impact how long the analysis will take to complete. We thus recommend exploring this before proceeding.

6 Expressed region-level analysis

6.1 Via regionMatrix()

Now that we have the data, we can identify expressed regions (ERs) by using a cutoff of 30 on the base-level mean coverage from these 12 samples. Once the regions have been identified, we can calculate a coverage matrix with one row per ER and one column per sample (12 in this case). For doing this calculation we need to know the length of the sequence reads, which in this study were 76 bp long.

Note that for this type of analysis there is no major benefit of filtering the data. Although it can be done if needed.

## Use regionMatrix()
system.time(regionMat <- regionMatrix(fullCov, cutoff = 30, L = 76))
## By using totalMapped equal to targetSize, regionMatrix() assumes that you have normalized the data already in fullCoverage(), loadCoverage() or filterData().
## 2017-08-21 19:56:47 regionMatrix: processing chr21
## 2017-08-21 19:56:47 filterData: normalizing coverage
## 2017-08-21 19:56:47 filterData: done normalizing coverage
## 2017-08-21 19:56:48 filterData: originally there were 48129895 rows, now there are 2256 rows. Meaning that 100 percent was filtered.
## 2017-08-21 19:56:48 findRegions: identifying potential segments
## 2017-08-21 19:56:48 findRegions: segmenting information
## 2017-08-21 19:56:48 findRegions: identifying candidate regions
## 2017-08-21 19:56:49 findRegions: identifying region clusters
## 2017-08-21 19:56:49 getRegionCoverage: processing chr21
## 2017-08-21 19:56:49 getRegionCoverage: done processing chr21
## 2017-08-21 19:56:49 regionMatrix: calculating coverageMatrix
## 2017-08-21 19:56:49 regionMatrix: adjusting coverageMatrix for 'L'
##    user  system elapsed 
##   2.116   0.048   2.166
## Explore results
class(regionMat)
## [1] "list"
names(regionMat$chr21)
## [1] "regions"        "coverageMatrix" "bpCoverage"

regionMatrix() returns a list of elements of length equal to the number of chromosomes analyzed. For each chromosome, there are three pieces of output. The actual ERs are arranged in a GRanges object named regions.

## regions output
regionMat$chr21$regions
## GRanges object with 45 ranges and 6 metadata columns:
##      seqnames               ranges strand |            value
##         <Rle>            <IRanges>  <Rle> |        <numeric>
##    1    chr21 [ 9827018,  9827582]      * | 313.671741938907
##    2    chr21 [15457301, 15457438]      * | 215.084607424943
##    3    chr21 [20230140, 20230192]      * | 38.8325471608144
##    4    chr21 [20230445, 20230505]      * | 41.3245082031834
##    5    chr21 [27253318, 27253543]      * | 34.9131305372469
##   ..      ...                  ...    ... .              ...
##   41    chr21 [33039644, 33039688]      * | 34.4705371008979
##   42    chr21 [33040784, 33040798]      * |  32.134222178989
##   43    chr21 [33040890, 33040890]      * | 30.0925002098083
##   44    chr21 [33040900, 33040901]      * | 30.1208333174388
##   45    chr21 [48019401, 48019414]      * | 31.1489284609755
##                  area indexStart  indexEnd cluster clusterL
##             <numeric>  <integer> <integer>   <Rle>    <Rle>
##    1 177224.534195483          1       565       1      565
##    2 29681.6758246422        566       703       2      138
##    3 2058.12499952316        704       756       3      366
##    4 2520.79500039419        757       817       3      366
##    5  7890.3675014178        818      1043       4      765
##   ..              ...        ...       ...     ...      ...
##   41 1551.17416954041       2180      2224      17       45
##   42 482.013332684835       2225      2239      18      118
##   43 30.0925002098083       2240      2240      18      118
##   44 60.2416666348775       2241      2242      18      118
##   45 436.084998453657       2243      2256      19       14
##   -------
##   seqinfo: 1 sequence from an unspecified genome
## Number of regions
length(regionMat$chr21$regions)
## [1] 45

bpCoverage is the base-level coverage list which can then be used for plotting.

## Base-level coverage matrices for each of the regions
## Useful for plotting
lapply(regionMat$chr21$bpCoverage[1:2], head, n = 2)
## $`1`
##   HSB113 HSB123 HSB126 HSB130 HSB135 HSB136 HSB145 HSB153 HSB159 HSB178
## 1  93.20   3.32  28.22   5.62 185.17  98.34   5.88  16.71   3.52  15.71
## 2 124.76   7.25  63.68  11.32 374.85 199.28  10.39  30.53   5.83  29.35
##   HSB92 HSB97
## 1 47.40 36.54
## 2 65.04 51.42
## 
## $`2`
##     HSB113 HSB123 HSB126 HSB130 HSB135 HSB136 HSB145 HSB153 HSB159 HSB178
## 566  45.59   7.94  15.92  34.75 141.61 104.21  19.87  38.61   4.97   23.2
## 567  45.59   7.94  15.92  35.15 141.64 104.30  19.87  38.65   4.97   23.2
##     HSB92 HSB97
## 566 13.95 22.21
## 567 13.95 22.21
## Check dimensions. First region is 565 long, second one is 138 bp long.
## The columns match the number of samples (12 in this case).
lapply(regionMat$chr21$bpCoverage[1:2], dim)
## $`1`
## [1] 565  12
## 
## $`2`
## [1] 138  12

The end result of the coverage matrix is shown below. Note that the coverage has been adjusted for read length. Because reads might not fully align inside a given region, the numbers are generally not integers but can be rounded if needed.

## Dimensions of the coverage matrix
dim(regionMat$chr21$coverageMatrix)
## [1] 45 12
## Coverage for each region. This matrix can then be used with limma or other pkgs
head(regionMat$chr21$coverageMatrix)
##         HSB113     HSB123      HSB126      HSB130      HSB135      HSB136
## 1 3653.1093346 277.072106 1397.068687 1106.722895 8987.460401 5570.221054
## 2  333.3740816  99.987237  463.909476  267.354342 1198.713552 1162.313418
## 3   35.3828948  20.153553   30.725394   23.483947   16.786842   17.168947
## 4   42.3398681  29.931579   41.094474   24.724736   32.634080   19.309606
## 5   77.7402631 168.939342  115.059342  171.861974  180.638684   93.503158
## 6    0.7988158   1.770263    1.473421    2.231053    1.697368    1.007895
##        HSB145       HSB153     HSB159      HSB178       HSB92        HSB97
## 1 1330.158818 1461.2986829 297.939342 1407.288552 1168.519079 1325.9622371
## 2  257.114210  313.8513139  67.940131  193.695657  127.543553  200.7834228
## 3   22.895921   52.8756585  28.145395   33.127368   23.758816   20.4623685
## 4   33.802632   51.6146040  31.244343   33.576974   29.546183   28.2011836
## 5   90.950526   36.3046051  78.069605   97.151316  100.085790   35.5428946
## 6    1.171316    0.4221053   1.000132    1.139079    1.136447    0.3956579

We can then use the coverage matrix and packages such as limma, DESeq2 or edgeR to identify which ERs are differentially expressed.

6.1.1 Find DERs with DESeq2

Here we’ll use DESeq2 to identify the DERs. To use it we need to round the coverage data.

## Required
library('DESeq2')

## Round matrix
counts <- round(regionMat$chr21$coverageMatrix)

## Round matrix and specify design
dse <- DESeqDataSetFromMatrix(counts, pheno, ~ group + gender)
## converting counts to integer mode
## Perform DE analysis
dse <- DESeq(dse, test = 'LRT', reduced = ~ gender, fitType = 'local')
## estimating size factors
## estimating dispersions
## gene-wise dispersion estimates
## mean-dispersion relationship
## final dispersion estimates
## fitting model and testing
## Extract results
deseq <- regionMat$chr21$regions
mcols(deseq) <- c(mcols(deseq), results(dse))

## Explore the results
deseq
## GRanges object with 45 ranges and 12 metadata columns:
##      seqnames               ranges strand |            value
##         <Rle>            <IRanges>  <Rle> |        <numeric>
##    1    chr21 [ 9827018,  9827582]      * | 313.671741938907
##    2    chr21 [15457301, 15457438]      * | 215.084607424943
##    3    chr21 [20230140, 20230192]      * | 38.8325471608144
##    4    chr21 [20230445, 20230505]      * | 41.3245082031834
##    5    chr21 [27253318, 27253543]      * | 34.9131305372469
##   ..      ...                  ...    ... .              ...
##   41    chr21 [33039644, 33039688]      * | 34.4705371008979
##   42    chr21 [33040784, 33040798]      * |  32.134222178989
##   43    chr21 [33040890, 33040890]      * | 30.0925002098083
##   44    chr21 [33040900, 33040901]      * | 30.1208333174388
##   45    chr21 [48019401, 48019414]      * | 31.1489284609755
##                  area indexStart  indexEnd cluster clusterL
##             <numeric>  <integer> <integer>   <Rle>    <Rle>
##    1 177224.534195483          1       565       1      565
##    2 29681.6758246422        566       703       2      138
##    3 2058.12499952316        704       756       3      366
##    4 2520.79500039419        757       817       3      366
##    5  7890.3675014178        818      1043       4      765
##   ..              ...        ...       ...     ...      ...
##   41 1551.17416954041       2180      2224      17       45
##   42 482.013332684835       2225      2239      18      118
##   43 30.0925002098083       2240      2240      18      118
##   44 60.2416666348775       2241      2242      18      118
##   45 436.084998453657       2243      2256      19       14
##               baseMean     log2FoldChange             lfcSE
##              <numeric>          <numeric>         <numeric>
##    1  2846.28716155756  -1.69031816472548  0.83195870366948
##    2  451.519643002767  -1.16404264400518 0.757489962635838
##    3  29.5780895864663 0.0461487746156367 0.458096633387309
##    4  36.0602688380417 -0.186620018986545 0.390920460570197
##    5  101.646763062388 -0.138737745290958 0.320166382651629
##   ..               ...                ...               ...
##   41  20.7820346002786 -0.642055952701049   0.4276611764306
##   42  6.41054227728292 -0.634320864108628  0.51226189777078
##   43 0.129716561731455 -0.859548956942508  3.11653965011299
##   44 0.702291488002413 -0.628284603274914  2.24737800886765
##   45  5.29329263119445  -1.69456340588639  1.25228952015329
##                    stat             pvalue              padj
##               <numeric>          <numeric>         <numeric>
##    1  0.215261662919431  0.642674256598179 0.997155030203294
##    2   0.87112644422632  0.350643649949685 0.997155030203294
##    3   3.13208182520646 0.0767656530295047 0.863613596581927
##    4   2.22570758632982  0.135730461463207 0.997155030203294
##    5   3.95798720528117 0.0466494518014198 0.862039631046612
##   ..                ...                ...               ...
##   41  0.604781356173675   0.43675951534617 0.997155030203294
##   42  0.545403917207146  0.460201756145951 0.997155030203294
##   43 0.0206273206873746  0.885798852352319 0.997155030203294
##   44  0.582510548110136  0.445329945344265 0.997155030203294
##   45   5.78959103593262 0.0161213391338242 0.725460261022089
##   -------
##   seqinfo: 1 sequence from an unspecified genome

You can get similar results using edgeR using these functions: DGEList(), calcNormFactors(), estimateGLMRobustDisp(), glmFit(), and glmLRT().

6.1.2 Find DERs with limma

Alternatively, we can find DERs using limma. Here we’ll exemplify a type of test closer to what we’ll do later with the F-statistics approach. First of all, we need to define our models.

## Build models
mod <- model.matrix(~ pheno$group + pheno$gender)
mod0 <- model.matrix(~ pheno$gender)

Next, we’ll transform the coverage information using the same default transformation from analyzeChr().

## Transform coverage
transformedCov <- log2(regionMat$chr21$coverageMatrix + 32)

We can then fit the models and get the F-statistic p-values and control the FDR.

## Example using limma
library('limma')
## 
## Attaching package: 'limma'
## The following object is masked from 'package:DESeq2':
## 
##     plotMA
## The following object is masked from 'package:BiocGenerics':
## 
##     plotMA
## Run limma
fit <- lmFit(transformedCov, mod)
fit0 <- lmFit(transformedCov, mod0)

## Determine DE status for the regions
## Also in https://github.com/LieberInstitute/jaffelab with help and examples
getF  <- function(fit, fit0, theData) {
    rss1 = rowSums((fitted(fit)-theData)^2)
    df1 = ncol(fit$coef)
    rss0 = rowSums((fitted(fit0)-theData)^2)
    df0 = ncol(fit0$coef)
    fstat = ((rss0-rss1)/(df1-df0))/(rss1/(ncol(theData)-df1))
    f_pval = pf(fstat, df1-df0, ncol(theData)-df1,lower.tail=FALSE)
    fout = cbind(fstat,df1-1,ncol(theData)-df1,f_pval)
    colnames(fout)[2:3] = c("df1","df0")
    fout = data.frame(fout)
    return(fout)
}

ff <- getF(fit, fit0, transformedCov)

## Get the p-value and assign it to the regions
limma <- regionMat$chr21$regions
limma$fstat <- ff$fstat
limma$pvalue <- ff$f_pval
limma$padj <- p.adjust(ff$f_pval, 'BH')

## Explore the results
limma
## GRanges object with 45 ranges and 9 metadata columns:
##      seqnames               ranges strand |            value
##         <Rle>            <IRanges>  <Rle> |        <numeric>
##    1    chr21 [ 9827018,  9827582]      * | 313.671741938907
##    2    chr21 [15457301, 15457438]      * | 215.084607424943
##    3    chr21 [20230140, 20230192]      * | 38.8325471608144
##    4    chr21 [20230445, 20230505]      * | 41.3245082031834
##    5    chr21 [27253318, 27253543]      * | 34.9131305372469
##   ..      ...                  ...    ... .              ...
##   41    chr21 [33039644, 33039688]      * | 34.4705371008979
##   42    chr21 [33040784, 33040798]      * |  32.134222178989
##   43    chr21 [33040890, 33040890]      * | 30.0925002098083
##   44    chr21 [33040900, 33040901]      * | 30.1208333174388
##   45    chr21 [48019401, 48019414]      * | 31.1489284609755
##                  area indexStart  indexEnd cluster clusterL
##             <numeric>  <integer> <integer>   <Rle>    <Rle>
##    1 177224.534195483          1       565       1      565
##    2 29681.6758246422        566       703       2      138
##    3 2058.12499952316        704       756       3      366
##    4 2520.79500039419        757       817       3      366
##    5  7890.3675014178        818      1043       4      765
##   ..              ...        ...       ...     ...      ...
##   41 1551.17416954041       2180      2224      17       45
##   42 482.013332684835       2225      2239      18      118
##   43 30.0925002098083       2240      2240      18      118
##   44 60.2416666348775       2241      2242      18      118
##   45 436.084998453657       2243      2256      19       14
##                  fstat             pvalue              padj
##              <numeric>          <numeric>         <numeric>
##    1  1.63845506978229  0.232544646826758 0.581361617066895
##    2  4.30744267104787 0.0677643558689958 0.324601395828103
##    3  1.32334173439853  0.279640574300262  0.62919129217559
##    4 0.380331651796345    0.5527043658146  0.86307385076635
##    5  7.24951917288678 0.0246954639590833 0.309531992076185
##   ..               ...                ...               ...
##   41  3.11798620713961  0.111243958775485 0.385075241915142
##   42  3.66183754305905 0.0879543238722494 0.329828714520935
##   43    3.878602667104 0.0804175220210141 0.328980771904149
##   44  4.39338051333778 0.0655381330160829 0.324601395828103
##   45  6.80914666972956 0.0282970022097315 0.309531992076185
##   -------
##   seqinfo: 1 sequence from an unspecified genome

In this simple example, none of the ERs have strong differential expression signal when adjusting for an FDR of 5%.

table(limma$padj < 0.05, deseq$padj < 0.05)
##        
##         FALSE
##   FALSE    45

6.2 Via railMatrix()

If you have Rail output, you can get the same results faster than with regionMatrix(). Rail will create the summarized coverage BigWig file for you, but we are not including it in this package due to its size. So, lets create it.

## Calculate the mean: this step takes a long time with many samples
meanCov <- Reduce('+', fullCov$chr21) / ncol(fullCov$chr21)

## Save it on a bigwig file called meanChr21.bw
createBw(list('chr21' = DataFrame('meanChr21' = meanCov)), keepGR = 
    FALSE)
## 2017-08-21 19:56:52 coerceGR: coercing sample meanChr21
## 2017-08-21 19:56:53 createBwSample: exporting bw for sample meanChr21

Now that we have the files Rail creates for us, we can use railMatrix().

## Identify files to use
summaryFile <- 'meanChr21.bw'
## We had already found the sample BigWig files and saved it in the object 'files'
## Lets just rename it to sampleFiles for clarity.
sampleFiles <- files

## Get the regions
system.time( 
    regionMat.rail <- railMatrix(chrs = 'chr21', summaryFiles = summaryFile, 
        sampleFiles = sampleFiles, L = 76, cutoff = 30, maxClusterGap = 3000L)
)
## 2017-08-21 19:56:55 loadCoverage: finding chromosome lengths
## 2017-08-21 19:56:55 loadCoverage: loading BigWig file meanChr21.bw
## 2017-08-21 19:56:55 loadCoverage: applying the cutoff to the merged data
## 2017-08-21 19:56:55 filterData: originally there were 48129895 rows, now there are 48129895 rows. Meaning that 0 percent was filtered.
## 2017-08-21 19:56:55 filterData: originally there were 48129895 rows, now there are 2256 rows. Meaning that 100 percent was filtered.
## 2017-08-21 19:56:55 findRegions: identifying potential segments
## 2017-08-21 19:56:55 findRegions: segmenting information
## 2017-08-21 19:56:55 .getSegmentsRle: segmenting with cutoff(s) 30
## 2017-08-21 19:56:55 findRegions: identifying candidate regions
## 2017-08-21 19:56:56 findRegions: identifying region clusters
## 2017-08-21 19:56:56 railMatrix: processing regions 1 to 45
## 2017-08-21 19:56:56 railMatrix: processing file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB113.bw
## 2017-08-21 19:56:56 railMatrix: processing file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB123.bw
## 2017-08-21 19:56:56 railMatrix: processing file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB126.bw
## 2017-08-21 19:56:56 railMatrix: processing file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB130.bw
## 2017-08-21 19:56:56 railMatrix: processing file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB135.bw
## 2017-08-21 19:56:56 railMatrix: processing file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB136.bw
## 2017-08-21 19:56:56 railMatrix: processing file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB145.bw
## 2017-08-21 19:56:57 railMatrix: processing file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB153.bw
## 2017-08-21 19:56:57 railMatrix: processing file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB159.bw
## 2017-08-21 19:56:57 railMatrix: processing file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB178.bw
## 2017-08-21 19:56:57 railMatrix: processing file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB92.bw
## 2017-08-21 19:56:57 railMatrix: processing file /home/biocbuild/bbs-3.5-bioc/R/library/derfinderData/extdata/AMY/HSB97.bw
##    user  system elapsed 
##   2.420   0.036   2.457

When you take into account the time needed to load the data (fullCoverage()) and then creating the matrix (regionMatrix()), railMatrix() is faster and less memory intensive.

The objects are not identical due to small rounding errors, but it’s nothing to worry about.

## Overall not identical due to small rounding errors
identical(regionMat, regionMat.rail)
## [1] FALSE
## Actual regions are the same
identical(ranges(regionMat$chr21$regions), ranges(regionMat.rail$chr21$regions))
## [1] TRUE
## When you round, the small differences go away
identical(round(regionMat$chr21$regions$value, 4),
    round(regionMat.rail$chr21$regions$value, 4))
## [1] TRUE
identical(round(regionMat$chr21$regions$area, 4),
    round(regionMat.rail$chr21$regions$area, 4))
## [1] TRUE

7 Single base-level F-statistics analysis

One form of base-level differential expression analysis implemented in derfinder is to calculate F-statistics for every base and use them to define candidate differentially expressed regions. This type of analysis is further explained in this section.

7.1 Models

Once we have the base-level coverage data for all 12 samples, we can construct the models. In this case, we want to find differences between fetal and adult samples while adjusting for gender and a proxy of the library size.

We can use sampleDepth() and it’s helper function collapseFullCoverage() to get a proxy of the library size. Note that you would normally use the unfiltered data from all the chromosomes in this step and not just one.

## Get some idea of the library sizes
sampleDepths <- sampleDepth(collapseFullCoverage(fullCov), 1)
## 2017-08-21 19:56:57 sampleDepth: Calculating sample quantiles
## 2017-08-21 19:56:57 sampleDepth: Calculating sample adjustments
sampleDepths
## HSB113.100% HSB123.100% HSB126.100% HSB130.100% HSB135.100% HSB136.100% 
##    19.82106    19.40505    19.53045    19.52017    20.33392    19.97758 
## HSB145.100% HSB153.100% HSB159.100% HSB178.100%  HSB92.100%  HSB97.100% 
##    19.49827    19.41285    19.24186    19.44252    19.55904    19.47733

sampleDepth() is similar to calcNormFactors() from metagenomeSeq with some code underneath tailored for the type of data we are using. collapseFullCoverage() is only needed to deal with the size of the data.

We can then define the nested models we want to use using makeModels(). This is a helper function that assumes that you will always adjust for the library size. You then need to define the variable to test, in this case we are comparing fetal vs adult samples. Optionally, you can adjust for other sample covariates, such as the gender in this case.

## Define models
models <- makeModels(sampleDepths, testvars = pheno$group,
    adjustvars = pheno[, c('gender')]) 

## Explore the models
lapply(models, head)
## $mod
##   (Intercept) testvarsadult sampleDepths adjustVar1M
## 1           1             0     19.82106           0
## 2           1             0     19.40505           1
## 3           1             0     19.53045           1
## 4           1             0     19.52017           1
## 5           1             0     20.33392           0
## 6           1             0     19.97758           0
## 
## $mod0
##   (Intercept) sampleDepths adjustVar1M
## 1           1     19.82106           0
## 2           1     19.40505           1
## 3           1     19.53045           1
## 4           1     19.52017           1
## 5           1     20.33392           0
## 6           1     19.97758           0

Note how the null model (mod0) is nested in the alternative model (mod). Use the same models for all your chromosomes unless you have a specific reason to use chromosome-specific models. Note that derfinder is very flexible and works with any type of nested model.

7.2 Find candidate DERs

Next, we can find candidate differentially expressed regions (DERs) using as input the segments of the genome where at least one sample has coverage greater than 2. That is, the filtered coverage version we created previously.

The main function in derfinder for this type of analysis is analyzeChr(). It works at a chromosome level and runs behinds the scenes several other derfinder functions. To use it, you have to provide the models, the grouping information, how to calculate the F-statistic cutoff and most importantly, the number of permutations.

By default analyzeChr() will use a theoretical cutoff. In this example, we use the cutoff that would correspond to a p-value of 0.05. To assign p-values to the candidate DERs, derfinder permutes the rows of the model matrices, re-calculates the F-statistics and identifies null regions. Then it compares the area of the observed regions versus the areas from the null regions to assign an empirical p-value.

In this example we will use twenty permutations, although in a real case scenario you might consider a larger number of permutations.

In real scenario, you might consider saving the results from all the chromosomes in a given directory. Here we will use analysisResults. For each chromosome you analyze, a new directory with the chromosome-specific data will be created. So in this case, we will have analysisResults/chr21.

## Create a analysis directory
dir.create('analysisResults')
originalWd <- getwd()
setwd(file.path(originalWd, 'analysisResults'))

## Perform differential expression analysis
system.time(results <- analyzeChr(chr = 'chr21', filteredCov$chr21, models,
    groupInfo = pheno$group, writeOutput = TRUE, cutoffFstat = 5e-02,
    nPermute = 20, seeds = 20140923 + seq_len(20), returnOutput = TRUE))
## 2017-08-21 19:56:58 analyzeChr: Pre-processing the coverage data
## 2017-08-21 19:57:01 analyzeChr: Calculating statistics
## 2017-08-21 19:57:01 calculateStats: calculating the F-statistics
## 2017-08-21 19:57:01 analyzeChr: Calculating pvalues
## 2017-08-21 19:57:01 analyzeChr: Using the following theoretical cutoff for the F-statistics 5.31765507157871
## 2017-08-21 19:57:01 calculatePvalues: identifying data segments
## 2017-08-21 19:57:01 findRegions: segmenting information
## 2017-08-21 19:57:01 findRegions: identifying candidate regions
## 2017-08-21 19:57:01 findRegions: identifying region clusters
## 2017-08-21 19:57:01 calculatePvalues: calculating F-statistics for permutation 1 and seed 20140924
## 2017-08-21 19:57:02 findRegions: segmenting information
## 2017-08-21 19:57:02 findRegions: identifying candidate regions
## 2017-08-21 19:57:02 calculatePvalues: calculating F-statistics for permutation 2 and seed 20140925
## 2017-08-21 19:57:02 findRegions: segmenting information
## 2017-08-21 19:57:02 findRegions: identifying candidate regions
## 2017-08-21 19:57:02 calculatePvalues: calculating F-statistics for permutation 3 and seed 20140926
## 2017-08-21 19:57:02 findRegions: segmenting information
## 2017-08-21 19:57:02 findRegions: identifying candidate regions
## 2017-08-21 19:57:02 calculatePvalues: calculating F-statistics for permutation 4 and seed 20140927
## 2017-08-21 19:57:02 findRegions: segmenting information
## 2017-08-21 19:57:02 findRegions: identifying candidate regions
## 2017-08-21 19:57:02 calculatePvalues: calculating F-statistics for permutation 5 and seed 20140928
## 2017-08-21 19:57:03 findRegions: segmenting information
## 2017-08-21 19:57:03 findRegions: identifying candidate regions
## 2017-08-21 19:57:03 calculatePvalues: calculating F-statistics for permutation 6 and seed 20140929
## 2017-08-21 19:57:04 findRegions: segmenting information
## 2017-08-21 19:57:04 findRegions: identifying candidate regions
## 2017-08-21 19:57:04 calculatePvalues: calculating F-statistics for permutation 7 and seed 20140930
## 2017-08-21 19:57:04 findRegions: segmenting information
## 2017-08-21 19:57:04 findRegions: identifying candidate regions
## 2017-08-21 19:57:04 calculatePvalues: calculating F-statistics for permutation 8 and seed 20140931
## 2017-08-21 19:57:04 findRegions: segmenting information
## 2017-08-21 19:57:04 findRegions: identifying candidate regions
## 2017-08-21 19:57:04 calculatePvalues: calculating F-statistics for permutation 9 and seed 20140932
## 2017-08-21 19:57:04 findRegions: segmenting information
## 2017-08-21 19:57:04 findRegions: identifying candidate regions
## 2017-08-21 19:57:04 calculatePvalues: calculating F-statistics for permutation 10 and seed 20140933
## 2017-08-21 19:57:05 findRegions: segmenting information
## 2017-08-21 19:57:05 findRegions: identifying candidate regions
## 2017-08-21 19:57:05 calculatePvalues: calculating F-statistics for permutation 11 and seed 20140934
## 2017-08-21 19:57:05 findRegions: segmenting information
## 2017-08-21 19:57:06 findRegions: identifying candidate regions
## 2017-08-21 19:57:06 calculatePvalues: calculating F-statistics for permutation 12 and seed 20140935
## 2017-08-21 19:57:06 findRegions: segmenting information
## 2017-08-21 19:57:06 findRegions: identifying candidate regions
## 2017-08-21 19:57:06 calculatePvalues: calculating F-statistics for permutation 13 and seed 20140936
## 2017-08-21 19:57:06 findRegions: segmenting information
## 2017-08-21 19:57:06 findRegions: identifying candidate regions
## 2017-08-21 19:57:06 calculatePvalues: calculating F-statistics for permutation 14 and seed 20140937
## 2017-08-21 19:57:06 findRegions: segmenting information
## 2017-08-21 19:57:06 findRegions: identifying candidate regions
## 2017-08-21 19:57:06 calculatePvalues: calculating F-statistics for permutation 15 and seed 20140938
## 2017-08-21 19:57:07 findRegions: segmenting information
## 2017-08-21 19:57:07 findRegions: identifying candidate regions
## 2017-08-21 19:57:07 calculatePvalues: calculating F-statistics for permutation 16 and seed 20140939
## 2017-08-21 19:57:07 findRegions: segmenting information
## 2017-08-21 19:57:07 findRegions: identifying candidate regions
## 2017-08-21 19:57:07 calculatePvalues: calculating F-statistics for permutation 17 and seed 20140940
## 2017-08-21 19:57:08 findRegions: segmenting information
## 2017-08-21 19:57:08 findRegions: identifying candidate regions
## 2017-08-21 19:57:08 calculatePvalues: calculating F-statistics for permutation 18 and seed 20140941
## 2017-08-21 19:57:08 findRegions: segmenting information
## 2017-08-21 19:57:08 findRegions: identifying candidate regions
## 2017-08-21 19:57:08 calculatePvalues: calculating F-statistics for permutation 19 and seed 20140942
## 2017-08-21 19:57:08 findRegions: segmenting information
## 2017-08-21 19:57:08 findRegions: identifying candidate regions
## 2017-08-21 19:57:08 calculatePvalues: calculating F-statistics for permutation 20 and seed 20140943
## 2017-08-21 19:57:09 findRegions: segmenting information
## 2017-08-21 19:57:09 findRegions: identifying candidate regions
## 2017-08-21 19:57:09 calculatePvalues: calculating the p-values
## 2017-08-21 19:57:09 analyzeChr: Annotating regions
## No annotationPackage supplied. Trying org.Hs.eg.db.
## Getting TSS and TSE.
## Getting CSS and CSE.
## Getting exons.
## Annotating genes.
## .....
##    user  system elapsed 
##  70.148   0.124  70.333

To speed up analyzeChr(), you might need to use several cores via the mc.cores argument. If memory is limiting, you might want to use a smaller chunksize (default is 5 million). Note that if you use too many cores, you might hit the input/output ceiling of your data network and/or hard drives speed.

Before running with a large number of permutations we recommend exploring how long each permutation cycle takes using a single permutation.

Note that analyzing each chromosome with a large number of permutations and a rich data set can take several hours, so we recommend running one job running analyzeChr() per chromosome, and then merging the results via mergeResults(). This process is further described in the advanced derfinder vignette.

7.3 Explore results

When using returnOutput = TRUE, analyzeChr() will return a list with the results to explore interactively. However, by default it writes the results to disk (one .Rdata file per result).

The following code explores the results.

## Explore
names(results)
## [1] "timeinfo"     "optionsStats" "coveragePrep" "fstats"      
## [5] "regions"      "annotation"

7.3.1 optionStats

optionStats stores the main options used in the analyzeChr() call including the models used, the type of cutoff, number of permutations, seeds for the permutations. All this information can be useful to reproduce the analysis.

## Explore optionsStats
names(results$optionsStats)
##  [1] "models"          "cutoffPre"       "scalefac"       
##  [4] "chunksize"       "cutoffFstat"     "cutoffType"     
##  [7] "nPermute"        "seeds"           "groupInfo"      
## [10] "lowMemDir"       "analyzeCall"     "cutoffFstatUsed"
## [13] "smooth"          "smoothFunction"  "weights"        
## [16] "returnOutput"
## Call used
results$optionsStats$analyzeCall
## analyzeChr(chr = "chr21", coverageInfo = filteredCov$chr21, models = models, 
##     cutoffFstat = 0.05, nPermute = 20, seeds = 20140923 + seq_len(20), 
##     groupInfo = pheno$group, writeOutput = TRUE, returnOutput = TRUE)

7.3.2 coveragePrep

coveragePrep has the result from the preprocessCoverage() step. This includes the genomic position index, the mean coverage (after scaling and the log2 transformation) for all the samples, and the group mean coverages. By default, the chunks are written to disk in optionsStats$lowMemDir (chr21/chunksDir in this example) to help reduce the required memory resources. Otherwise it is stored in coveragePrep$coverageProcessed.

## Explore coveragePrep
names(results$coveragePrep)
## [1] "coverageProcessed" "mclapplyIndex"     "position"         
## [4] "meanCoverage"      "groupMeans"
## Group means
results$coveragePrep$groupMeans
## $fetal
## numeric-Rle of length 130356 with 116452 runs
##   Lengths:                 1                 1 ...                 1
##   Values : 0.401666664828857 0.428333345800638 ...  1.24833332498868
## 
## $adult
## numeric-Rle of length 130356 with 119226 runs
##   Lengths:                 1                 1 ...                 1
##   Values : 0.406666670615474  0.41333334085842 ...   1.6266666551431

7.3.3 fstats

The F-statistics are then stored in fstats. These are calculated using calculateStats().

## Explore optionsStats
results$fstats
## numeric-Rle of length 130356 with 126807 runs
##   Lengths:                   1                   1 ...                   1
##   Values :  0.0192260952849003  0.0299693665742238 ...    2.32463844414503
## Note that the length matches the number of bases used
identical(length(results$fstats), sum(results$coveragePrep$position))
## [1] TRUE

7.3.4 regions

The candidate DERs and summary results from the permutations is then stored in regions. This is the output from calculatePvalues() which uses several underneath other functions including calculateStats() and findRegions().

## Explore regions
names(results$regions)
## [1] "regions"         "nullStats"       "nullWidths"      "nullPermutation"

For the null regions, the summary information is composed of the mean F-statistic for the null regions (regions$nullStats), the width of the null regions (regions$nullWidths), and the permutation number under which they were identified (regions$nullPermutation).

## Permutation summary information
results$regions[2:4]
## $nullStats
## numeric-Rle of length 16738 with 16736 runs
##   Lengths:                1                1 ...                1
##   Values : 12.2254943226418 5.42502250701367 ... 9.56804518038913
## 
## $nullWidths
## integer-Rle of length 16738 with 14779 runs
##   Lengths:   1   1   1   1   1   1   1   1 ...   1   1   1   1   1   1   1
##   Values :   8   2  49   1   2   1   2   1 ...   7   3  25   2  20   1 151
## 
## $nullPermutation
## integer-Rle of length 16738 with 20 runs
##   Lengths: 1128 3248  858 1472  406 1512 ...  912  296  266  196  330 1104
##   Values :    1    2    3    4    5    6 ...   15   16   17   18   19   20

The most important part of the output is the GRanges object with the candidate DERs shown below.

## Candidate DERs
results$regions$regions
## GRanges object with 591 ranges and 14 metadata columns:
##      seqnames               ranges strand |     value      area indexStart
##         <Rle>            <IRanges>  <Rle> | <numeric> <numeric>  <integer>
##   up    chr21 [47610386, 47610682]      * | 11.103042  3297.603     122158
##   up    chr21 [40196145, 40196444]      * | 10.061425  3018.427      76110
##   up    chr21 [27253616, 27253948]      * |  8.434883  2808.816      22019
##   up    chr21 [22115534, 22115894]      * |  7.236451  2612.359      12274
##   up    chr21 [22914853, 22915064]      * |  9.780659  2073.500      17318
##   ..      ...                  ...    ... .       ...       ...        ...
##   up    chr21 [35889784, 35889784]      * |  5.319523  5.319523      60088
##   up    chr21 [47610093, 47610093]      * |  5.319119  5.319119     121865
##   up    chr21 [16333728, 16333728]      * |  5.318807  5.318807       5048
##   up    chr21 [34001896, 34001896]      * |  5.318713  5.318713      32577
##   up    chr21 [34809571, 34809571]      * |  5.318014  5.318014      43694
##       indexEnd cluster clusterL meanCoverage meanfetal  meanadult
##      <integer>   <Rle>    <Rle>    <numeric> <numeric>  <numeric>
##   up    122454     138      933    1.5979517  0.822890  2.3730135
##   up     76409      71     1323    1.3035083  2.025322  0.5816944
##   up     22351      28      407   33.6578579 42.467042 24.8486737
##   up     12634       9      694    0.9644645  1.719058  0.2098707
##   up     17529      21      217    2.8389780  4.235928  1.4420283
##   ..       ...     ...      ...          ...       ...        ...
##   up     60088      51      742     2.754167  3.360000   2.148333
##   up    121865     138      933     1.455833  0.775000   2.136667
##   up      5048       1        9     1.195000  1.231667   1.158333
##   up     32577      38     1428     1.712500  2.333333   1.091667
##   up     43694      46      149     2.950000  2.878333   3.021667
##      log2FoldChangeadultvsfetal    pvalues significant   qvalues
##                       <numeric>  <numeric>    <factor> <numeric>
##   up                  1.5279488 0.00985722        TRUE 0.9484048
##   up                 -1.7998180 0.01129100        TRUE 0.9484048
##   up                 -0.7731748 0.01224685        TRUE 0.9484048
##   up                 -3.0340455 0.01296374        TRUE 0.9484048
##   up                 -1.5545785 0.01834040        TRUE 0.9484048
##   ..                        ...        ...         ...       ...
##   up                -0.64524334  0.9979688       FALSE 0.9996416
##   up                 1.46309367  0.9983273       FALSE 0.9996416
##   up                -0.08856135  0.9988052       FALSE 0.9996416
##   up                -1.09586004  0.9991636       FALSE 0.9996416
##   up                 0.07011082  0.9996416       FALSE 0.9996416
##      significantQval
##             <factor>
##   up           FALSE
##   up           FALSE
##   up           FALSE
##   up           FALSE
##   up           FALSE
##   ..             ...
##   up           FALSE
##   up           FALSE
##   up           FALSE
##   up           FALSE
##   up           FALSE
##   -------
##   seqinfo: 1 sequence from an unspecified genome; no seqlengths

The metadata columns are:

  • value is the mean F-statistics for the candidate DER.
  • area is the sum of the F-statistics for the candidate DER.
  • indexStart Relates the genomic start coordinate with the filtered genomic index start coordinate.
  • indexEnd Similarly as above but for the end coordinates.
  • cluster The cluster id to which this candidate DER belongs to.
  • clusterL The length of the cluster to which this candidate DER belongs to.
  • meanCoverage The base level mean coverage for the candidate DER.
  • meanfetal In this example, the mean coverage for the fetal samples.
  • meanadult In this example, the mean coverage for the adult samples.
  • log2FoldChangeadultvsfetal In this example, the log2 fold change between adult vs fetal samples.
  • pvalues The p-value for the candidate DER.
  • significant By default, whether the p-value is less than 0.05 or not.
  • qvalues The q-value for the candidate DER calculated with qvalue.
  • significantQval By default, whether the q-value is less than 0.10 or not.

Note that for this type of analysis you might want to try a few coverage cutoffs and/or F-statistic cutoffs. One quick way to evaluate the results is to compare the width of the regions.

## Width of potential DERs
summary(width(results$regions$regions))
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##    1.00    1.00    4.00   17.98   17.00  361.00
sum(width(results$regions$regions) > 50)
## [1] 68
## Width of candidate DERs
sig <- as.logical(results$regions$regions$significant)
summary(width(results$regions$regions[ sig ]))
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##    76.0    97.5   115.0   163.4   193.5   361.0
sum(width(results$regions$regions[ sig ]) > 50)
## [1] 19

7.3.5 Nearest annotation

analyzeChr() will find the nearest annotation feature using matchGenes() from bumphunter (version >= 1.7.3). This information is useful considering that the candidate DERs were identified without relying on annotation. Yet at the end, we are interested to check if they are inside a known exon, upstream a gene, etc.

## Nearest annotation
head(results$annotation)
##        name
## 1       LSS
## 2      ETS2
## 3       APP
## 4 LINC00320
## 5     NCAM2
## 6 LINC00320
##                                                                                                                                                                                                                                                                                                                                                              annotation
## 1                                           NM_001001438 NM_001145436 NM_001145437 NM_002340 NP_001001438 NP_001138908 NP_001138909 NP_002331 XM_006724004 XM_011529564 XM_011529565 XM_011529566 XM_011529567 XM_017028346 XM_017028347 XM_017028348 XP_006724067 XP_011527866 XP_011527867 XP_011527868 XP_011527869 XP_016883835 XP_016883836 XP_016883837 XR_937491
## 2                                                                                                                                                                                                                                                                     NM_001256295 NM_005239 NP_001243224 NP_005230 XM_005260935 XM_017028290 XP_005260992 XP_016883779
## 3                                                                                                                     NM_000484 NM_001136016 NM_001136129 NM_001136130 NM_001136131 NM_001204301 NM_001204302 NM_001204303 NM_201413 NM_201414 NP_000475 NP_001129488 NP_001129601 NP_001129602 NP_001129603 NP_001191230 NP_001191231 NP_001191232 NP_958816 NP_958817
## 4                                                                                                                                                                                                                                                                                                                               NR_024090 NR_109786 NR_109787 NR_109788
## 5 NM_004540 NP_004531 XM_011529575 XM_011529576 XM_011529579 XM_011529580 XM_011529581 XM_011529582 XM_011529585 XM_017028353 XM_017028354 XM_017028355 XM_017028356 XM_017028357 XM_017028358 XP_011527877 XP_011527878 XP_011527881 XP_011527882 XP_011527883 XP_011527884 XP_011527887 XP_016883842 XP_016883843 XP_016883844 XP_016883845 XP_016883846 XP_016883847
## 6                                                                                                                                                                                                                                                                                                                               NR_024090 NR_109786 NR_109787 NR_109788
##   description     region distance   subregion insideDistance exonnumber
## 1 inside exon     inside    38056 inside exon              0         22
## 2 inside exon     inside    18390 inside exon              0         10
## 3 inside exon     inside   289498 inside exon              0         16
## 4 inside exon     inside    59532 inside exon              0          7
## 5  downstream downstream   544220        <NA>             NA         NA
## 6 inside exon     inside    59315 inside exon              0          7
##   nexons                         UTR strand  geneL codingL Entrez
## 1     22                       3'UTR      -  39700      NA   4047
## 2     10                       3'UTR      +  19123      NA   2114
## 3     16                       3'UTR      - 290585      NA    351
## 4      7 inside transcription region      -  60513      NA 387486
## 5     15                        <NA>      + 541884      NA   4685
## 6      7 inside transcription region      -  60513      NA 387486
##   subjectHits
## 1         421
## 2         203
## 3         354
## 4         407
## 5         435
## 6         407

For more details on the output please check the bumphunter package.

Check the section about non-human data (specifically, using annotation different from hg19) on the advanced vignette.

7.3.6 Time spent

The final piece is the wallclock time spent during each of the steps in analyzeChr().

## Time spent
results$timeinfo
##                      init                     setup 
## "2017-08-21 19:56:58 EDT" "2017-08-21 19:56:58 EDT" 
##                  prepData                  savePrep 
## "2017-08-21 19:57:00 EDT" "2017-08-21 19:57:01 EDT" 
##            calculateStats                 saveStats 
## "2017-08-21 19:57:01 EDT" "2017-08-21 19:57:01 EDT" 
##             saveStatsOpts          calculatePvalues 
## "2017-08-21 19:57:01 EDT" "2017-08-21 19:57:09 EDT" 
##                  saveRegs                  annotate 
## "2017-08-21 19:57:09 EDT" "2017-08-21 19:58:08 EDT" 
##                  saveAnno 
## "2017-08-21 19:58:08 EDT"
## Use this information to make a plot
timed <- diff(results$timeinfo)
timed.df <- data.frame(Seconds = as.numeric(timed), Step = factor(names(timed),
    levels = rev(names(timed))))
library('ggplot2')
ggplot(timed.df, aes(y = Step, x = Seconds)) + geom_point()

7.4 Merge results

Once you have analyzed each chromosome using analyzeChr(), you can use mergeResults() to merge the results. This function does not return an object in R but instead creates several Rdata files with the main results from the different chromosomes.

## Go back to the original directory
setwd(originalWd)

## Merge results from several chromosomes. In this case we only have one.
mergeResults(chrs='chr21', prefix="analysisResults",
    genomicState = genomicState$fullGenome, 
    optionsStats = results$optionsStats)
## 2017-08-21 19:58:10 mergeResults: Saving options used
## 2017-08-21 19:58:10 Loading chromosome chr21
## 2017-08-21 19:58:10 mergeResults: calculating FWER
## 2017-08-21 19:58:10 mergeResults: Saving fullNullSummary
## 2017-08-21 19:58:10 mergeResults: Re-calculating the p-values
## 2017-08-21 19:58:10 mergeResults: Saving fullRegions
## 2017-08-21 19:58:10 mergeResults: assigning genomic states
## 2017-08-21 19:58:10 annotateRegions: counting
## 2017-08-21 19:58:10 annotateRegions: annotating
## 2017-08-21 19:58:10 mergeResults: Saving fullAnnotatedRegions
## 2017-08-21 19:58:10 mergeResults: Saving fullFstats
## 2017-08-21 19:58:10 mergeResults: Saving fullTime
## Files created by mergeResults()
dir('analysisResults', pattern = '.Rdata')
## [1] "fullAnnotatedRegions.Rdata" "fullFstats.Rdata"          
## [3] "fullNullSummary.Rdata"      "fullRegions.Rdata"         
## [5] "fullTime.Rdata"             "optionsMerge.Rdata"

7.4.1 optionsMerge

For reproducibility purposes, the options used the merge the results are stored in optionsMerge.

## Options used to merge
load(file.path('analysisResults', 'optionsMerge.Rdata'))

## Contents
names(optionsMerge)
## [1] "chrs"            "significantCut"  "minoverlap"      "mergeCall"      
## [5] "cutoffFstatUsed" "optionsStats"
## Merge call
optionsMerge$mergeCall
## mergeResults(chrs = "chr21", prefix = "analysisResults", genomicState = genomicState$fullGenome, 
##     optionsStats = results$optionsStats)

7.4.2 fullRegions

The main result from mergeResults() is in fullRegions. This is a GRanges object with the candidate DERs from all the chromosomes. It also includes the nearest annotation metadata as well as FWER adjusted p-values (fwer) and whether the FWER adjusted p-value is less than 0.05 (significantFWER).

## Load all the regions
load(file.path('analysisResults', 'fullRegions.Rdata'))

## Metadata columns
names(mcols(fullRegions))
##  [1] "value"                      "area"                      
##  [3] "indexStart"                 "indexEnd"                  
##  [5] "cluster"                    "clusterL"                  
##  [7] "meanCoverage"               "meanfetal"                 
##  [9] "meanadult"                  "log2FoldChangeadultvsfetal"
## [11] "pvalues"                    "significant"               
## [13] "qvalues"                    "significantQval"           
## [15] "name"                       "annotation"                
## [17] "description"                "region"                    
## [19] "distance"                   "subregion"                 
## [21] "insideDistance"             "exonnumber"                
## [23] "nexons"                     "UTR"                       
## [25] "annoStrand"                 "geneL"                     
## [27] "codingL"                    "Entrez"                    
## [29] "subjectHits"                "fwer"                      
## [31] "significantFWER"

Note that analyzeChr() only has the information for a given chromosome at a time, so mergeResults() re-calculates the p-values and q-values using the information from all the chromosomes.

7.4.3 fullAnnotatedRegions

In preparation for visually exploring the results, mergeResults() will run annotateRegions() which counts how many known exons, introns and intergenic segments each candidate DER overlaps (by default with a minimum overlap of 20bp). annotateRegions() uses a summarized version of the genome annotation created with makeGenomicState(). For this example, we can use the data included in derfinder which is the summarized annotation of hg19 for chromosome 21.

## Load annotateRegions() output
load(file.path('analysisResults', 'fullAnnotatedRegions.Rdata'))

## Information stored
names(fullAnnotatedRegions)
## [1] "countTable"     "annotationList"
## Take a peak
lapply(fullAnnotatedRegions, head)
## $countTable
##   exon intergenic intron
## 1    1          0      0
## 2    1          0      0
## 3    1          0      0
## 4    1          0      0
## 5    0          1      0
## 6    1          0      0
## 
## $annotationList
## GRangesList object of length 6:
## $1 
## GRanges object with 1 range and 4 metadata columns:
##        seqnames               ranges strand |   theRegion
##           <Rle>            <IRanges>  <Rle> | <character>
##   4871    chr21 [47609038, 47611149]      - |        exon
##                        tx_id                              tx_name
##                <IntegerList>                      <CharacterList>
##   4871 73448,73449,73450,... uc002zij.3,uc002zik.2,uc002zil.2,...
##                 gene
##        <IntegerList>
##   4871           170
## 
## $2 
## GRanges object with 1 range and 4 metadata columns:
##        seqnames               ranges strand | theRegion       tx_id
##   1189    chr21 [40194598, 40196878]      + |      exon 72757,72758
##                      tx_name gene
##   1189 uc002yxf.3,uc002yxg.4   77
## 
## $3 
## GRanges object with 1 range and 4 metadata columns:
##        seqnames               ranges strand | theRegion
##   2965    chr21 [27252861, 27254082]      - |      exon
##                        tx_id                              tx_name gene
##   2965 73066,73067,73068,... uc002ylz.3,uc002yma.3,uc002ymb.3,...  139
## 
## ...
## <3 more elements>
## -------
## seqinfo: 1 sequence from hg19 genome

8 ChIP-seq differential binding

As of version 1.5.27 derfinder has parameters that allow smoothing of the single base-level F-statistics before determining DERs. This allows finding differentially bounded regions (peaks) using ChIP-seq data. In general, ChIP-seq studies are smaller than RNA-seq studies which means that the single base-level F-statistics approach is well suited for differential binding analysis.

To smooth the F-statistics use smooth = TRUE in analyzeChr(). The default smoothing function is bumphunter::locfitByCluster() and all its parameters can be passed specified in the call to analyzeChr(). In particular, the minNum and bpSpan arguments are important. We recommend setting minNum to the minimum read length and bpSpan to the average peak length expected in the ChIP-seq data being analyzed. Smoothing the F-statistics will take longer but not use significantly more memory than the default behavior. So take this into account when choosing the number of permutations to run.

9 Visually explore results

Optionally, we can use the addon package derfinderPlot to visually explore the results.

To make the region level plots, we will need to extract the region level coverage data. We can do so using getRegionCoverage() as shown below.

## Find overlaps between regions and summarized genomic annotation
annoRegs <- annotateRegions(fullRegions, genomicState$fullGenome)
## 2017-08-21 19:58:11 annotateRegions: counting
## 2017-08-21 19:58:11 annotateRegions: annotating
## Indeed, the result is the same because we only used chr21
identical(annoRegs, fullAnnotatedRegions)
## [1] FALSE
## Get the region coverage
regionCov <- getRegionCoverage(fullCov, fullRegions)
## 2017-08-21 19:58:11 getRegionCoverage: processing chr21
## 2017-08-21 19:58:11 getRegionCoverage: done processing chr21
## Explore the result
head(regionCov[[1]])
##   HSB113 HSB123 HSB126 HSB130 HSB135 HSB136 HSB145 HSB153 HSB159 HSB178
## 1   0.68   0.44   0.48   0.36   0.19   2.34   1.29   1.77   2.21   2.69
## 2   0.60   0.44   0.48   0.36   0.19   2.30   1.29   1.77   2.21   2.64
## 3   0.60   0.40   0.48   0.32   0.19   2.39   1.37   1.81   2.31   2.69
## 4   0.64   0.40   0.48   0.32   0.19   2.61   1.42   1.89   2.36   2.88
## 5   0.64   0.40   0.48   0.36   0.19   2.65   1.42   1.93   2.36   2.88
## 6   0.60   0.44   0.48   0.39   0.23   2.65   1.59   1.93   2.36   2.83
##   HSB92 HSB97
## 1  1.89  3.57
## 2  1.86  3.60
## 3  1.89  3.60
## 4  1.89  3.60
## 5  1.96  3.70
## 6  1.93  3.70

With this, we are all set to visually explore the results.

library('derfinderPlot')

## Overview of the candidate DERs in the genome
plotOverview(regions = fullRegions, annotation = results$annotation,
    type = 'fwer')

suppressPackageStartupMessages(library('TxDb.Hsapiens.UCSC.hg19.knownGene'))
txdb <- TxDb.Hsapiens.UCSC.hg19.knownGene

## Base-levle coverage plots for the first 10 regions
plotRegionCoverage(regions = fullRegions, regionCoverage = regionCov, 
    groupInfo = pheno$group, nearestAnnotation = results$annotation, 
    annotatedRegions = annoRegs, whichRegions=1:10, txdb = txdb, scalefac = 1, 
    ask = FALSE)

## Cluster plot for the first region
plotCluster(idx = 1, regions = fullRegions, annotation = results$annotation,
    coverageInfo = fullCov$chr21, txdb = txdb, groupInfo = pheno$group,
    titleUse = 'fwer')

The quick start to using derfinder has example plots for the expressed regions-level approach. The vignette for derfinderPlot has even more examples.

10 Interactive HTML reports

We have also developed an addon package called regionReport available via Bioconductor.

The function derfinderReport() in regionReport basically takes advantage of the results from mergeResults() and plotting functions available in derfinderPlot as well as other neat features from knitrBootstrap (with contributions from Andrew J. Bass, Dabney, and Robinson, 2015). It then generates a customized report for single-base level F-statistics DER finding analyses.

For results from regionMatrix() or railMatrix() use renderReport() from regionReport. In both cases, the resulting HTML report promotes reproducibility of the analysis and allows you to explore in more detail the results through some diagnostic plots.

We think that these reports are very important when you are exploring the resulting DERs after changing a key parameter in analyzeChr(), regionMatrix() or railMatrix().

Check out the vignette for regionReport for example reports generated with it.

11 Miscellaneous features

In this section we go over some other features of derfinder which can be useful for performing feature-counts based analyses, exploring the results, or exporting data.

11.1 Feature level analysis

Similar to the expressed region-level analysis, you might be interested in performing a feature-level analysis. More specifically, this means getting a count matrix at the exon-level (or gene-level). coverageToExon() allows you to get such a matrix by taking advantage of the summarized annotation produced by makeGenomicState().

In this example, we use the genomic state included in the package which has the information for chr21 TxDb.Hsapiens.UCSC.hg19.knownGene annotation.

## Get the exon-level matrix
system.time(exonCov <- coverageToExon(fullCov, genomicState$fullGenome, L = 76))
## class: SerialParam
##   bpisup: TRUE; bpnworkers: 1; bptasks: 0; bpjobname: BPJOB
##   bplog: FALSE; bpthreshold: INFO; bpstopOnError: TRUE
##   bptimeout: 2592000; bpprogressbar: FALSE
##   bplogdir: NA
## class: SerialParam
##   bpisup: TRUE; bpnworkers: 1; bptasks: 0; bpjobname: BPJOB
##   bplog: FALSE; bpthreshold: INFO; bpstopOnError: TRUE
##   bptimeout: 2592000; bpprogressbar: FALSE
##   bplogdir: NA
## 2017-08-21 19:58:16 coverageToExon: processing chromosome chr21
## class: SerialParam
##   bpisup: TRUE; bpnworkers: 1; bptasks: 0; bpjobname: BPJOB
##   bplog: FALSE; bpthreshold: INFO; bpstopOnError: TRUE
##   bptimeout: 2592000; bpprogressbar: FALSE
##   bplogdir: NA
## 2017-08-21 19:58:21 coverageToExon: processing chromosome chr21
##    user  system elapsed 
##   9.464   2.796  12.291
## Dimensions of the matrix
dim(exonCov)
## [1] 2658   12
## Explore a little bit
tail(exonCov)
##         HSB113      HSB123      HSB126      HSB130       HSB135
## 4983 0.1173684   0.1382895   0.0000000   0.0000000   0.07894737
## 4985 4.6542105   1.4557895   1.6034210   0.7798684   1.29592103
## 4986 7.1510526 291.0205261 252.4892106 152.1905258 439.69500020
## 4988 0.0000000   0.7063158   0.7223684   0.2639474   0.48657895
## 4990 2.3064474  64.9423686  69.6584212  35.5769737 101.76394746
## 4992 0.1652632   8.1834211   9.7538158   2.4193421  10.08618420
##            HSB136       HSB145     HSB153      HSB159    HSB178
## 4983   0.03947368 5.263158e-03  0.1052632  0.04934211 0.1289474
## 4985   0.72986842 9.964474e-01 10.4906578  4.39013167 7.3119736
## 4986 263.63486853 2.345414e+02  4.5310526 10.64368415 4.9807895
## 4988   0.37657895 6.156579e-01  0.0000000  0.00000000 0.0000000
## 4990  76.96736838 5.978842e+01  1.2284210  2.67486840 0.9542105
## 4992  14.41013157 5.325658e+00  0.2402632  0.73039474 0.2601316
##            HSB92       HSB97
## 4983  0.03986842  0.44605264
## 4985  1.65184211 10.87723678
## 4986 11.40447366  6.23315790
## 4988  0.05644737  0.00000000
## 4990  3.98013159  1.96131579
## 4992  0.66907895  0.07473684

With this matrix, rounded if necessary, you can proceed to use packages such as limma, DESeq2, edgeR among others.

11.2 Compare results visually

We can certainly make region-level plots using plotRegionCoverage() or cluster plots using plotCluster() or overview plots using plotOveview(), all from derfinderPlot.

First we need to get the relevant annotation information.

## Annotate regions as exonic, intronic or intergenic
system.time(annoGenome <- annotateRegions(regionMat$chr21$regions,
    genomicState$fullGenome))
## 2017-08-21 19:58:25 annotateRegions: counting
## 2017-08-21 19:58:25 annotateRegions: annotating
##    user  system elapsed 
##   0.420   0.004   0.424
## Note that the genomicState object included in derfinder only has information
## for chr21 (hg19).

## Identify closest genes to regions
suppressPackageStartupMessages(library('bumphunter'))
suppressPackageStartupMessages(library('TxDb.Hsapiens.UCSC.hg19.knownGene'))
txdb <- TxDb.Hsapiens.UCSC.hg19.knownGene
genes <- annotateTranscripts(txdb)
## No annotationPackage supplied. Trying org.Hs.eg.db.
## Getting TSS and TSE.
## Getting CSS and CSE.
## Getting exons.
## Annotating genes.
system.time(annoNear <- matchGenes(regionMat$chr21$regions, genes))
##    user  system elapsed 
##   2.060   0.000   2.062

Now we can proceed to use derfinderPlot to make the region-level plots for the top 100 regions.

## Identify the top regions by highest total coverage
top <- order(regionMat$chr21$regions$area, decreasing = TRUE)[1:100]

## Base-level plots for the top 100 regions with transcript information
library('derfinderPlot')
plotRegionCoverage(regionMat$chr21$regions,
    regionCoverage = regionMat$chr21$bpCoverage,
    groupInfo = pheno$group, nearestAnnotation = annoNear,
    annotatedRegions = annoGenome, whichRegions = top, scalefac = 1,
    txdb = txdb, ask = FALSE)

However, we can alternatively use epivizr to view the candidate DERs and the region matrix results in a genome browser.

## Load epivizr, it's available from Bioconductor
library('epivizr')

## Load data to your browser
mgr <- startEpiviz()
ders_dev <- mgr$addDevice(
    fullRegions[as.logical(fullRegions$significantFWER) ], "Candidate DERs")
ders_potential_dev <- mgr$addDevice(
    fullRegions[!as.logical(fullRegions$significantFWER) ], "Potential DERs")
regs_dev <- mgr$addDevice(regionMat$chr21$regions, "Region Matrix")

## Go to a place you like in the genome
mgr$navigate("chr21", start(regionMat$chr21$regions[top[1]]) - 100,
    end(regionMat$chr21$regions[top[1]]) + 100)

## Stop the navigation
mgr$stopServer()

11.3 Export coverage to BigWig files

derfinder also includes createBw() with related functions createBwSample() and coerceGR() to export the output of fullCoverage() to BigWig files. These functions can be useful in the case where you start with BAM files and later on want to save the coverage data into BigWig files, which are generally smaller.

## Subset only the first sample
fullCovSmall <- lapply(fullCov, '[', 1)

## Export to BigWig
bw <- createBw(fullCovSmall)
## 2017-08-21 19:58:58 coerceGR: coercing sample HSB113
## 2017-08-21 19:58:58 createBwSample: exporting bw for sample HSB113
## See the file. Note that the sample name is used to name the file.
dir(pattern = '.bw')
## [1] "HSB113.bw"    "meanChr21.bw"
## Internally createBw() coerces each sample to a GRanges object before 
## exporting to a BigWig file. If more than one sample was exported, the
## GRangesList would have more elements.
bw
## GRangesList object of length 1:
## $HSB113 
## GRanges object with 155950 ranges and 1 metadata column:
##         seqnames               ranges strand |              score
##            <Rle>            <IRanges>  <Rle> |          <numeric>
##   chr21    chr21   [9458667, 9458741]      * | 0.0399999991059303
##   chr21    chr21   [9540957, 9540971]      * | 0.0399999991059303
##   chr21    chr21   [9543719, 9543778]      * | 0.0399999991059303
##   chr21    chr21   [9651480, 9651554]      * | 0.0399999991059303
##   chr21    chr21   [9653397, 9653471]      * | 0.0399999991059303
##     ...      ...                  ...    ... .                ...
##   chr21    chr21 [48093246, 48093255]      * | 0.0399999991059303
##   chr21    chr21 [48093257, 48093331]      * | 0.0399999991059303
##   chr21    chr21 [48093350, 48093424]      * | 0.0399999991059303
##   chr21    chr21 [48112194, 48112268]      * | 0.0399999991059303
##   chr21    chr21 [48115056, 48115130]      * | 0.0399999991059303
## 
## -------
## seqinfo: 1 sequence from an unspecified genome

12 Advanced arguments

If you are interested in using advanced arguments in derfinder, they are described in the manual pages of each function. Some of the most common advanced arguments are:

verbose controls whether to print status updates for nearly all the functions. chrsStyle is used to determine the chromosome naming style and is powered by GenomeInfoDb. Note that chrsStyle is used in any of the functions that call extendedMapSeqlevels(). If you are working with a different organism than Homo sapiens set the global species option using options(species = 'your species') with the syntax used in names(GenomeInfoDb::genomeStyles()). If you want to disable extendedMapSeqlevels() set chrsStyle to NULL, which can be useful if your organism is not part of GenomeInfoDb.

The third commonly used advanced argument is mc.cores. It controls the number of cores to use for the functions that can run with more than one core to speed up. In nearly all the cases, the maximum number of cores depends on the number of chromosomes. One notable exception is analyzeChr() where the maximum number of cores depends on the chunksize used and the dimensions of the data for the chromosome under study.

Note that using the ... argument allows you to specify some of the documented arguments. For example, you might want to control the maxClusterGap from findRegions() in the analyzeChr() call.

13 Non-human data

If you are working with data from an organism that is not Homo sapiens, then set the global options defining the species and the chrsStyle used. For example, if you are working with Arabidopsis Thaliana and the NCBI naming style, then set the options using the following code:

## Set global species and chrsStyle options
options(species = 'arabidopsis_thaliana')
options(chrsStyle = 'NCBI')

## Then proceed to load and analyze the data

Internally derfinder uses extendedMapSeqlevels() to use the appropriate chromosome naming style given a species in all functions involving chromosome names.

Further note that the argument txdb from analyzeChr() is passed to bumphunter::annotateTranscripts(txdb). So if you are using a genome different from hg19 remember to provide the appropriate annotation data or simply use analyzeChr(runAnnotation = FALSE).

So, in the Arabidopsis Thaliana example, your analyzeChr() call would look like this:

## Load transcript database information
library('TxDb.Athaliana.BioMart.plantsmart28')

## Set organism options
options(species = 'arabidopsis_thaliana')
options(chrsStyle = 'NCBI')

## Run command
analyzeChr(txdb = TxDb.Athaliana.BioMart.plantsmart28, --Other arguments--)

You might find the discussion Using bumphunter with non-human genomes useful.

14 Functions that use multiple cores

Currently, the following functions can use multiple cores, several of which are called inside analyzeChr().

All parallel operations use SnowParam() from BiocParallel when more than 1 core is being used. Otherwise, SerialParam() is used. Note that if you prefer to specify other types of parallelization you can do so by specifying the BPPARAM.custom advanced argument.

Because SnowParam() requires R to load the necessary packages on each worker, the key function fstats.apply() was isolated in the derfinderHelper package. This package has much faster loading speeds than derfinder which greatly impacts performance on cases where the actual step of calculating the F-statistics is fast.

You may prefer to use MulticoreParam() described in the BiocParallel vignette. In that case, when using these functions use BPPARAM.custom = MulticoreParam(workers = x) where x is the number of cores you want to use. Note that in some systems, as is the case of the cluster used by derfinder’s developers, the system tools for assessing memory usage can be misleading, thus resulting in much higher memory loads when using MulticoreParam() instead of the default SnowParam().

15 Loading data details

15.1 Controlling loading from BAM files

If you are loading data from BAM files, you might want to specify some criteria to decide which reads to include or not. For example, your data might have been generated by a strand-specific protocol. You can do so by specifying the arguments of scanBamFlag() from Rsamtools.

You can also control whether to include or exclude bases with CIGAR string D (deletion from the reference) by setting the advanced argument drop.D = TRUE in your fullCoverage() or loadCoverage() call.

15.2 Unfiltered base-level coverage

Note that in most scenarios, the fullCov object illustrated in the introductory vignette can be large in memory. When making plots or calculating the region-level coverage, we don’t need the full information. In such situations, it might pay off to create a smaller version by loading only the required data. This can be achieved using the advanced argument which to fullCoverage() or loadCoverage().

However, it is important to consider that when reading the data from BAM files, a read might align partially inside the region of interest. By default such a read would be discarded and thus the base-level coverage would be lower than what it is in reality. The advanced argument protectWhich extends regions by 30 kbp (15 kbp each side) to help mitigate this issue.

We can illustrate this issue with the example data from derfinder. First, we load in the data and generate some regions of interest.

## Find some regions to work with
example('loadCoverage', 'derfinder')
## 
## ldCvrg> datadir <- system.file('extdata', 'genomeData', package='derfinder')
## 
## ldCvrg> files <- rawFiles(datadir = datadir, samplepatt = '*accepted_hits.bam$', 
## ldCvrg+     fileterm = NULL)
## 
## ldCvrg> ## Shorten the column names
## ldCvrg> names(files) <- gsub('_accepted_hits.bam', '', names(files))
## 
## ldCvrg> ## Read and filter the data, only for 2 files
## ldCvrg> dataSmall <- loadCoverage(files = files[1:2], chr = '21', cutoff = 0)
## 2017-08-21 19:58:58 loadCoverage: finding chromosome lengths
## 2017-08-21 19:58:58 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009101_accepted_hits.bam
## 2017-08-21 19:58:58 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009102_accepted_hits.bam
## 2017-08-21 19:58:59 loadCoverage: applying the cutoff to the merged data
## 2017-08-21 19:58:59 filterData: originally there were 48129895 rows, now there are 454 rows. Meaning that 100 percent was filtered.
## 
## ldCvrg> ## Not run: 
## ldCvrg> ##D ## Export to BigWig files
## ldCvrg> ##D createBw(list('chr21' = dataSmall))
## ldCvrg> ##D 
## ldCvrg> ##D ## Load data from BigWig files
## ldCvrg> ##D dataSmall.bw <- loadCoverage(c(ERR009101 = 'ERR009101.bw', ERR009102 = 
## ldCvrg> ##D     'ERR009102.bw'), chr = 'chr21')
## ldCvrg> ##D 
## ldCvrg> ##D ## Compare them
## ldCvrg> ##D mapply(function(x, y) { x - y }, dataSmall$coverage, dataSmall.bw$coverage)
## ldCvrg> ##D 
## ldCvrg> ##D ## Note that the only difference is the type of Rle (integer vs numeric) used
## ldCvrg> ##D ## to store the data.
## ldCvrg> ##D 
## ldCvrg> ## End(Not run)
## ldCvrg> 
## ldCvrg> 
## ldCvrg>
example('getRegionCoverage', 'derfinder')
## 
## gtRgnC> ## Obtain fullCov object
## gtRgnC> fullCov <- list('21'=genomeDataRaw$coverage)
## 
## gtRgnC> ## Assign chr lengths using hg19 information, use only first two regions
## gtRgnC> library('GenomicRanges')
## 
## gtRgnC> data(hg19Ideogram, package = 'biovizBase', envir = environment())
## 
## gtRgnC> regions <- genomeRegions$regions[1:2]
## 
## gtRgnC> seqlengths(regions) <- seqlengths(hg19Ideogram)[names(seqlengths(regions))]
## 
## gtRgnC> ## Finally, get the region coverage
## gtRgnC> regionCov <- getRegionCoverage(fullCov=fullCov, regions=regions)
## extendedMapSeqlevels: sequence names mapped from NCBI to UCSC for species homo_sapiens
## 2017-08-21 19:58:59 getRegionCoverage: processing chr21
## 2017-08-21 19:58:59 getRegionCoverage: done processing chr21

Next, we load the coverage again using which but without any padding. We can see how the coverage is not the same by looking at the maximum coverage for each sample.

## Illustrate reading data from a set of regions
test <- loadCoverage(files = files, chr = '21', cutoff = NULL, which = regions,
    protectWhich = 0, fileStyle = 'NCBI')
## extendedMapSeqlevels: sequence names mapped from UCSC to NCBI for species homo_sapiens
## 2017-08-21 19:58:59 loadCoverage: finding chromosome lengths
## 2017-08-21 19:58:59 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009101_accepted_hits.bam
## 2017-08-21 19:58:59 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009102_accepted_hits.bam
## 2017-08-21 19:58:59 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009105_accepted_hits.bam
## 2017-08-21 19:58:59 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009107_accepted_hits.bam
## 2017-08-21 19:58:59 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009108_accepted_hits.bam
## 2017-08-21 19:58:59 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009112_accepted_hits.bam
## 2017-08-21 19:58:59 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009115_accepted_hits.bam
## 2017-08-21 19:58:59 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009116_accepted_hits.bam
## 2017-08-21 19:58:59 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009131_accepted_hits.bam
## 2017-08-21 19:58:59 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009138_accepted_hits.bam
## 2017-08-21 19:58:59 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009144_accepted_hits.bam
## 2017-08-21 19:58:59 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009145_accepted_hits.bam
## 2017-08-21 19:58:59 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009148_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009151_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009152_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009153_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009159_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009161_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009163_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009164_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009167_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031812_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031835_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031867_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031868_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031900_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031904_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031914_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031936_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031958_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031960_accepted_hits.bam
## 2017-08-21 19:59:00 loadCoverage: applying the cutoff to the merged data
## 2017-08-21 19:59:00 filterData: originally there were 48129895 rows, now there are 48129895 rows. Meaning that 0 percent was filtered.
## Some reads were ignored and thus the coverage is lower as can be seen below:
sapply(test$coverage, max) - sapply(genomeDataRaw$coverage, max)
## ERR009101 ERR009102 ERR009105 ERR009107 ERR009108 ERR009112 ERR009115 
##         0         0         0         0        -1         0        -1 
## ERR009116 ERR009131 ERR009138 ERR009144 ERR009145 ERR009148 ERR009151 
##        -2        -2        -2        -1        -1        -3        -3 
## ERR009152 ERR009153 ERR009159 ERR009161 ERR009163 ERR009164 ERR009167 
##         0        -3        -3        -3        -1        -3        -3 
## SRR031812 SRR031835 SRR031867 SRR031868 SRR031900 SRR031904 SRR031914 
##         0        -1         0         0         0        -1         0 
## SRR031936 SRR031958 SRR031960 
##         0         0         0

When we re-load the data using some padding to the regions, we find that the coverage matches at all the bases.

## Illustrate reading data from a set of regions

test2 <- loadCoverage(files = files, chr = '21', cutoff = NULL, 
    which = regions, protectWhich = 3e4, fileStyle = 'NCBI')
## extendedMapSeqlevels: sequence names mapped from UCSC to NCBI for species homo_sapiens
## 2017-08-21 19:59:01 loadCoverage: finding chromosome lengths
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009101_accepted_hits.bam
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009102_accepted_hits.bam
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009105_accepted_hits.bam
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009107_accepted_hits.bam
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009108_accepted_hits.bam
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009112_accepted_hits.bam
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009115_accepted_hits.bam
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009116_accepted_hits.bam
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009131_accepted_hits.bam
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009138_accepted_hits.bam
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009144_accepted_hits.bam
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009145_accepted_hits.bam
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009148_accepted_hits.bam
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009151_accepted_hits.bam
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009152_accepted_hits.bam
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009153_accepted_hits.bam
## 2017-08-21 19:59:01 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009159_accepted_hits.bam
## 2017-08-21 19:59:02 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009161_accepted_hits.bam
## 2017-08-21 19:59:02 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009163_accepted_hits.bam
## 2017-08-21 19:59:02 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009164_accepted_hits.bam
## 2017-08-21 19:59:02 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/ERR009167_accepted_hits.bam
## 2017-08-21 19:59:02 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031812_accepted_hits.bam
## 2017-08-21 19:59:02 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031835_accepted_hits.bam
## 2017-08-21 19:59:02 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031867_accepted_hits.bam
## 2017-08-21 19:59:02 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031868_accepted_hits.bam
## 2017-08-21 19:59:02 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031900_accepted_hits.bam
## 2017-08-21 19:59:02 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031904_accepted_hits.bam
## 2017-08-21 19:59:02 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031914_accepted_hits.bam
## 2017-08-21 19:59:02 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031936_accepted_hits.bam
## 2017-08-21 19:59:02 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031958_accepted_hits.bam
## 2017-08-21 19:59:02 loadCoverage: loading BAM file /tmp/RtmpeKEf1X/Rinst4cd2a727a6a/derfinder/extdata/genomeData/SRR031960_accepted_hits.bam
## 2017-08-21 19:59:02 loadCoverage: applying the cutoff to the merged data
## 2017-08-21 19:59:03 filterData: originally there were 48129895 rows, now there are 48129895 rows. Meaning that 0 percent was filtered.
## Adding some padding to the regions helps get the same coverage
identical(sapply(test2$coverage, max), sapply(genomeDataRaw$coverage, max))
## [1] TRUE
## A more detailed test reveals that the coverage matches at every base
all(mapply(function(x, y) { identical(x, y) }, test2$coverage, genomeDataRaw$coverage))
## [1] TRUE

How much padding you need to use will depend on your specific data set, and you might be comfortable getting approximately the same coverage values for the sake of greatly reducing the memory resources needed.

15.3 Input files in a different naming style

If you are under the case where you like to use a specific chromosome naming style but the raw data files use another one, you might need to use the fileStyle argument.

For example, you could be working with Homo sapiens data and your preferred naming style is UCSC (chr1, chr2, …, chrX, chrY) but the raw data uses NCBI style names (1, 2, …, X, Y). In that case, use fullCoverage(fileStyle = 'NCBI') or loadCoverage(fileStyle = 'NCBI') depending if you are loading one chromosome or multiple at a time.

15.4 Loading data in chunks

If you prefer to do so, fullCoverage() and loadCoverage() can load the data of a chromosome in chunks using GenomicFiles. This is controlled by whether you specify the tilewidth argument.

Notice that you might run into slight coverage errors near the borders of the tiles for the same reason that was illustrated previously when loading specific regions.

This approach is not necessarily more efficient and can be significantly time consuming if you use a small tilewidth.

15.5 Large number of samples

If you have a large number of samples (say thousands), it might be best to submit cluster jobs that run loadCoverage() or fullCoverage() for only one chromosome at a time.

In the case of working with Rail output, you can either use railMatrix() with the argument file.cores greater than 1 or specify the advanced argument BPPARAM.railChr to control the parallel environment used for loading the BigWig files. Doing so, you can then submit cluster jobs that run railMatrix() for one chromosome at a time, yet read in the data fast. You can do all of from R by using BPPARAM.custom and BPPARAM.railChr at the same time where you use a BatchJobsParam() for the first one.

Another option when working with Rail output is to simply load the summary BigWig data (mean, median), then define the ERs using findRegions() and write them to a BED file. You can then use bwtool to create the coverage matrix.

16 Flow charts

16.1 DER analysis flow chart

The following figure illustrates how most of derfinder’s functions interact when performing a base-level differential expression analysis by calculating base-level F-statistics.

DER flow

Flow chart of the different processing steps (black boxes) that can be carried out using derfinder and the functions that perform these actions (in red). Input and output is shown in green boxes. Functions in blue are those applied to the results from multiple chromosomes (mergeResults() and derfinderReport). regionReport functions are shown in orange while derfinderPlot functions are shown in dark purple. Purple dotted arrow marks functions that require unfiltered base-level coverage.

16.2 analyzeChr() flow chart