-
Notifications
You must be signed in to change notification settings - Fork 212
General deepTools FAQs
For instructions on using deepTools 2.0 or newer, please go here. This page only applies to deepTools 1.5
- How does deepTools handle data from paired-end sequencing?
- I just want to try out a tool, how can I optimize the computation time?
- Can I specify more than one chromosome in the --regions Option?
- When should I exclude regions from computeGCbias?
- When should I use bamCoverage, when bamCompare?
- How does computeMatrix handle overlapping genome regions?
- Why does the maximum value in the heatmap not equal the maximum value in the matrix?
- How can I increase the resolution of the heatmap?
- How can I change the automatic labels of the clusters in a k-means clustered heatmap?
- How can I manually specify several groups of regions (instead of clustering)?
- What do I have to pay attention to when working with a draft version of a genome?
- How do I calculate the effective genome size for an organism that's not in your list?
- Where can I download the 2bit genome files required for computeGCbias?
Generally, all the modules working with BAM files (bamCorrelate, bamCoverage, bamCompare, bamFingerprint, computeGCbias) recognize paired-end sequencing data. You can enforce to ignore the fragment length based on the mate pairs using the option doNotExtendPairedEnds ("advanced options" in Galaxy).
When you're playing around with the tools to see what kinds of results they will produce, you can limit the operation to one chromosome or a specific region save computation time. In Galaxy, you will find this under "advanced output options" → "Region of the genome to limit the operation to"; the command line option is called "--region" (CHR:START:END).
The following tools currently have this option:
It works as follows: first, the entire genome represented in the BAM file will be regarded and sampled, then all the regions or sampled bins that do not overlap with the region indicated by the user will be discarded.
Be aware that you can limit the operation to only one chromosome (or one specific locus on a chromosome), if you would like to limit the operation to more than one region, see the answer below.
Several programs have the option to limit the operation to a specific region (i.e. to reduce computation time). The input must be given in the format of chr:start:end, for example "chr10" or "chr10:456700:891000". It is not possible to indicate more than one region, e.g. chr10, chr11 - this will not work!
Here are some ideas for workarounds if you do find yourself in a situation where you would in fact like to limit the operation to more than one region:
- general workaround: since all the tools that have the --region option work on BAM files, you could filter your reads prior to running the program, e.g. using intersectBed with --abam or samtools view. Use the resulting (smaller) BAM file for the deepTools program of your choice.
samtools view -b -L regionsOfInterest.bed Reads.bam > ReadsOverlappingWithRegionsOfInterest.bam
intersectBed -abam Reads.bam -b regionsOfInterest.bed > ReadsOverlappingWithRegionsOfInterest.bam
computeGCbias and bamCorrelate offer in-build solutions:
- bamCorrelate: bamCorrelate has two modes, bins and BED. If you make use of the BED mode (for details, see here), you can supply a BED file of regions that you would like to limit the operation to. This will do the same thing as in the general workaround mentioned above.
- computeGCbias: You can make use of the --filterOut option of computeGCbias. You will first need to create a BED file that contains all the regions you are not interested in. Then supply this file of RegionsOf__Non__Interest.bed to computeGCbias.
In general, we recommend that you should only correct for GC bias (using computeGCbias followed by correctGCbias) if you observe that the majority of the genome (the region between 30-60%) is continuously GC-biased and you want to compare this sample with another sample that is not GC-biased.
Sometimes, a certain GC bias is expected, for example for ChIP samples of H3K4me3 in mammalian samples where GC-rich promoters are expected to be enriched. To not confound the GC bias caused by the library preparation with the inherent, expected GC bias, we incorporated the possibility to supply a file of regions to computeGCbias that will be excluded from the GC bias calculation. This file should typically contain those regions that one expects to be significantly enriched per se. This way, the computeGCbias will focus on background regions.
Both tools produce bigWig files, i.e. they translate the read-centered information from a BAM file into scores for genomic regions of a fixed size. The only difference is the number of BAM files that the tools use as input: while bamCoverage will only take one BAM file and produce a coverage file that is mostly normalized for sequencing depth, bamCompare will take two BAM files that can be compared with each other using several mathematical operations. bamCompare will always normalize for sequencing depth like bamCoverage, but then it will perform additional calculations depending on what the user chose, for example:
-
bamCompare:
- ChIP vs. [input][] → obtain a bigWig file of log2ratios(ChIP/input)
- treatment vs. control → obtain a bigWig file of differences (Treatment - control)
- Replicate 1 and Replicate 2 → obtain a bigWig file where the values from two BAM files are summed up
If the BED file supplied to computeMatrix contains regions that overlap, computeMatrix will report those regions and issue warnings, but they will just be taken as is. If you would like to avoid that overlapping regions are taken into account, you will need to clean the BED file prior to using computeMatrix. There are several possibilities for modifying your BED file. Let's say your file looks like this:
$ cat testBed.bed
chr1 10 20 region1
chr1 7 15 region2
chr1 18 29 region3
chr1 35 40 region4
chr1 10 20 region1Duplicate
- if you just want to eliminate identical entries (here: region1 and region1Duplicate), use sort and uniq in the shell (note that the label of the identical regions is different - as uniq can only ignore fields at the beginning of a file, use rev to revert the sorted file, then uniq with ignoring the first field (which is now the name column) and then revert back)
$ sort -k1,1 -k2,2n testBed.bed | rev | uniq -f1 | rev
chr1 10 20 region1
chr1 7 15 region2
chr1 18 29 region3
chr1 35 40 region4
- if you would like to merge all overlapping regions to one big one, use the BEDtool mergeBed
- again, the BED file must be sorted first
- -n and -nms tell mergeBed to output the number of overlapping regions and the names of them
- in the resulting file, regions 1, 2 and 3 are merged
$ sort -k1,1 -k2,2n testBed.bed | mergeBed -i stdin -n -nms
chr1 7 29 region2;region1;region1Duplicate;region3 4
chr1 35 40 region4 1
- if you would like to keep only regions that do not overlap with any other region in the same BED file, use the same mergeBed routine but subsequently filter out those regions where several regions were merged
- the awk command will check the last field of each line ($NF) and will print the original line ($0) only if the last field contained a number smaller than 2
$ sort -k1,1 -k2,2n testBed.bed | mergeBed -i stdin -n -nms | awk '$NF < 2 {print $0}'
chr1 35 40 region4 1
The entire question went like this:
The max scale in the heatmap does not correspond to the max. values in the matrix. This suggests that some scaling/normalization is done prior to heatmap plot. (ex: I have values in the matrix going up to 50 and the scales on the matrix goes from 0 to ~13)
Indeed, if you do not indicate --zMax and/or --zMin together with the heatmapper command, the default of heatmapper is to remove outliers. We have found that this is beneficial in most cases as outliers tend to screw up the color schemes. In the code, this looks like this:
if zMin is None:
matrixFlatten = flattenMatrix(hm.matrixDict)
# try to avoid outliers by using np.percentile
zMin = np.percentile(matrixFlatten, 1.0)
if np.isnan(zMin):
zMin = None
...
if zMax is None:
if matrixFlatten is None:
matrixFlatten = flattenMatrix(hm.matrixDict)
# try to avoid outliers by using np.percentile
zMax = np.percentile(matrixFlatten, 98.0)
if np.isnan(zMax):
zMax = None
If you would like to include the entire range, you can set --zMax and --zMin to the min and max values you find in your matrix.
- decrease the bin size when generating the matrix using computeMatrix
- go to "advanced options" → "Length, in base pairs, of the non-overlapping bin for averaging the score over the regions length" → define a smaller value, e.g. 50 or 25 bp
- make sure, however, that you used a sufficiently small bin size when calculating the bigWig file, though (if generated with deepTools, you can check the option "bin size")
Each cluster will get its own box, exactly the same way as different groups of regions. Therefore, you can use the same option to define the labels of the final heatmap: In Galaxy: Heatmapper → "Advanced output options" → "Labels for the regions plotted in the heatmap".
If you indicated 3 clusters for k-means clustering, enter here: C1, C2, C3 → instead of the full default label ("cluster 1"), the heatmap will be labeled with the abbreviations.
In the command line, use the --regionsLabel option to define your customized names.
If you would like to compare the profiles and heatmaps for different groups of regions that you specified yourself, you need to tell computeMatrix to assess the BED file of regions accordingly. In Galaxy, you should add one BED file per group (i.e. genes.bed, exons.bed, introns.bed) while on the command line you should have just one BED file total with hashed lines indicating the end of a group. For more details, see the step-by-step-description here.
If you are working with sequences from a genome that is not included in our standard descriptions, you need to pay attention to two sets of data that you might need using deepTools:
- Effective genome size - this is mostly needed for bamCoverage and bamCompare, see below for details
- Reference genome sequence in 2bit format - this is needed for computeGCbias, see below for details
We plan on including a module that will calculate the effective genome size for you, but it's not ready yet, so you will have to find a solution outside of deepTools for the time being.
The "real" effective genome size is the part of the genome that is uniquely mappable. This means that the value will depend on the genome properties (how many repetitive elements, quality of the assembly etc.) and the length of the sequenced reads as 100 million 36-bp-reads might cover less than 100 million 100-bp-reads.
We currently have these options for you:
2. Use faCount (only if you let reads be aligned non-uniquely, too!)
1. Use an external tool There is a tool that promises to calculate the mappability for any genome given the read length (k-mer length): GEM-Mappability Calculator. According to this reply here, you can calculate the effective genome size after running this program by counting the numbers of "!" which stands for uniquely mappable regions.
2. Use faCount If you are using bowtie2 which reports multi-read alignments (= including non-uniquely mapping reads) as a default setting, you can use faCount from UCSC tools to report the total number of bases as well as the number of bases that are missing from the genome assembly indicated by 'N'. The effective genome size would then be the total number of base pairs minus the total number of 'N'. Here's an example output of faCount on D. melanogaster genome version dm3:
$ UCSCtools/faCount dm3.fa
#seq len A C G T N cpg
chr2L 23011544 6699731 4811687 4815192 6684734 200 926264
chr2LHet 368872 90881 58504 57899 90588 71000 10958
chr2R 21146708 6007371 4576037 4574750 5988450 100 917644
chr2RHet 3288761 828553 537840 529242 826306 566820 99227
chr3L 24543557 7113242 5153576 5141498 7135141 100 995078
chr3LHet 2555491 725986 473888 479000 737434 139183 89647
chr3R 27905053 7979156 5995211 5980227 7950459 0 1186894
chr3RHet 2517507 678829 447155 446597 691725 253201 84175
chr4 1351857 430227 238155 242039 441336 100 43274
chrU 10049037 2511952 1672330 1672987 2510979 1680789 335241
chrUextra 29004656 7732998 5109465 5084891 7614402 3462900 986216
chrX 22422827 6409325 4742952 4748415 6432035 90100 959534
chrXHet 204112 61961 40017 41813 60321 0 754
chrYHet 347038 74566 45769 47582 74889 104232 8441
chrM 19517 8152 2003 1479 7883 0 132
total 168736537 47352930 33904589 33863611 47246682 6368725 6650479
In this example: Total no. bp = 168,736,537 Total no. 'N' = 6,368,725
NOTE: this method only works if multi-reads are mapped randomly to their possible locations, in other words if repetitive regions are covered by reads in which case the effective genome size is the size of the genome discarding stretches of 'N's.
3. Use bamCoverage If you have a sample where you expect the genome to be covered completely, e.g. from genome sequencing, a very trivial solution is to use bamCoverage with a bin size of 1 bp and the --outFileFormat option set to 'bedgraph'. You can then count the number of non-Zero bins (= bp) which will indicate the mappable genome size for this specific sample.
4. Use genomeCoverageBed The BEDtool genomeCoverageBed can be used to calculate the number of bp in the genome for which 0 reads can be found overlapping. As described on the BEDtools website, you need:
- a file with the choromosome sizes of your sample's organism
- a position-sorted BAM file
bedtools genomecov -ibam sortedBAMfile.bam -g genome.size
The 2bit files of most genomes can be found here. Search for the .2bit ending. Otherwise, fasta files can be converted to 2bit using the UCSC programm faToTwoBit (available for different platforms from here
[read]: https://github.com/fidelram/deepTools/wiki/Glossary#terminology "the DNA piece that was actually sequenced ("read") by the sequencing machine (usually between 30 to 100 bp long, depending on the read-length of the sequencing protocol)" [input]: https://github.com/fidelram/deepTools/wiki/Glossary#terminology "confusing, albeit commonly used name for the 'no-antibody' control sample for ChIP experiments"
deepTools is developed by the Bioinformatics Facility at the Max Planck Institute for Immunobiology and Epigenetics, Freiburg. For troubleshooting, see our FAQ and get in touch: [email protected]
Wiki Start Page | Code | deepTools Galaxy | FAQ | Glossary | Gallery |