Abstract
A basic task in the analysis of count data from RNA-seq is the detection of differentially expressed genes. The count data are presented as a table which reports, for each sample, the number of sequence fragments that have been assigned to each gene. Analogous data also arise for other assay types, including comparative ChIP-Seq, HiC, shRNA screening, and mass spectrometry. An important analysis question is the quantification and statistical inference of systematic changes between conditions, as compared to within-condition variability. The package DESeq2 provides methods to test for differential expression by use of negative binomial generalized linear models; the estimates of dispersion and logarithmic fold changes incorporate data-driven prior distributions. This vignette explains the use of the package and demonstrates typical workflows. An RNA-seq workflow on the Bioconductor website covers similar material to this vignette but at a slower pace, including the generation of count matrices from FASTQ files. DESeq2 package version: 1.44.0
Note: if you use DESeq2 in published research, please cite:
Love, M.I., Huber, W., Anders, S. (2014) Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biology, 15:550. 10.1186/s13059-014-0550-8
Other Bioconductor packages with similar aims are edgeR, limma, DSS, EBSeq, and baySeq.
Here we show the most basic steps for a differential expression analysis. There are a variety of steps upstream of DESeq2 that result in the generation of counts or estimated counts for each sample, which we will discuss in the sections below. This code chunk assumes that you have a count matrix called cts
and a table of sample information called coldata
. The design
indicates how to model the samples, here, that we want to measure the effect of the condition, controlling for batch differences. The two factor variables batch
and condition
should be columns of coldata
.
dds <- DESeqDataSetFromMatrix(countData = cts,
colData = coldata,
design= ~ batch + condition)
dds <- DESeq(dds)
resultsNames(dds) # lists the coefficients
res <- results(dds, name="condition_trt_vs_untrt")
# or to shrink log fold changes association with condition:
res <- lfcShrink(dds, coef="condition_trt_vs_untrt", type="apeglm")
The following starting functions will be explained below:
DESeqDataSetFromTximport()
.DESeqDataSet()
.DESeqDataSetFromHTSeq()
.Any and all DESeq2 questions should be posted to the Bioconductor support site, which serves as a searchable knowledge base of questions and answers:
https://support.bioconductor.org
Posting a question and tagging with “DESeq2” will automatically send an alert to the package authors to respond on the support site. See the first question in the list of Frequently Asked Questions (FAQ) for information about how to construct an informative post.
You should not email your question to the package authors, as we will just reply that the question should be posted to the Bioconductor support site.
Constantin Ahlmann-Eltze has contributed core code for increasing the computational performance of DESeq2 and building an interface to his glmGamPoi package.
We have benefited in the development of DESeq2 from the help and feedback of many individuals, including but not limited to:
The Bionconductor Core Team, Alejandro Reyes, Andrzej Oles, Aleksandra Pekowska, Felix Klein, Nikolaos Ignatiadis (IHW), Anqi Zhu (apeglm), Joseph Ibrahim (apeglm), Vince Carey, Owen Solberg, Ruping Sun, Devon Ryan, Steve Lianoglou, Jessica Larson, Christina Chaivorapol, Pan Du, Richard Bourgon, Willem Talloen, Elin Videvall, Hanneke van Deutekom, Todd Burwell, Jesse Rowley, Igor Dolgalev, Stephen Turner, Ryan C Thompson, Tyr Wiesner-Hanks, Konrad Rudolph, David Robinson, Mingxiang Teng, Mathias Lesche, Sonali Arora, Jordan Ramilowski, Ian Dworkin, Bjorn Gruning, Ryan McMinds, Paul Gordon, Leonardo Collado Torres, Enrico Ferrero, Peter Langfelder, Gavin Kelly, Rob Patro, Charlotte Soneson, Koen Van den Berge, Fanny Perraudeau, Davide Risso, Stephan Engelbrecht, Nicolas Alcala, Jeremy Simon, Travis Ptacek, Rory Kirchner, R. Butler, Ben Keith, Dan Liang, Nil Aygün, Rory Nolan, Michael Schubert, Hugo Tavares, Eric Davis, Wancen Mu, Zhang Cheng, Frederik Ziebell, Luca Menestrina, Hendrik Weisse, I-Hsuan Lin, Rasmus Henningsson.
DESeq2 and its developers have been partially supported by funding from the European Union’s 7th Framework Programme via Project RADIANT, NIH NHGRI R01-HG009937, and by a CZI EOSS award.
As input, the DESeq2 package expects count data as obtained, e.g., from RNA-seq or another high-throughput sequencing experiment, in the form of a matrix of integer values. The value in the i-th row and the j-th column of the matrix tells how many reads can be assigned to gene i in sample j. Analogously, for other types of assays, the rows of the matrix might correspond e.g. to binding regions (with ChIP-Seq) or peptide sequences (with quantitative mass spectrometry). We will list method for obtaining count matrices in sections below.
The values in the matrix should be un-normalized counts or estimated counts of sequencing reads (for single-end RNA-seq) or fragments (for paired-end RNA-seq). The RNA-seq workflow describes multiple techniques for preparing such count matrices. It is important to provide count matrices as input for DESeq2’s statistical model (Love, Huber, and Anders 2014) to hold, as only the count values allow assessing the measurement precision correctly. The DESeq2 model internally corrects for library size, so transformed or normalized values such as counts scaled by library size should not be used as input.
The object class used by the DESeq2 package to store the read counts and the intermediate estimated quantities during statistical analysis is the DESeqDataSet, which will usually be represented in the code here as an object dds
.
A technical detail is that the DESeqDataSet class extends the RangedSummarizedExperiment class of the SummarizedExperiment package. The “Ranged” part refers to the fact that the rows of the assay data (here, the counts) can be associated with genomic ranges (the exons of genes). This association facilitates downstream exploration of results, making use of other Bioconductor packages’ range-based functionality (e.g. find the closest ChIP-seq peaks to the differentially expressed genes).
A DESeqDataSet object must have an associated design formula. The design formula expresses the variables which will be used in modeling. The formula should be a tilde (~) followed by the variables with plus signs between them (it will be coerced into an formula if it is not already). The design can be changed later, however then all differential analysis steps should be repeated, as the design formula is used to estimate the dispersions and to estimate the log2 fold changes of the model.
Note: In order to benefit from the default settings of the package, you should put the variable of interest at the end of the formula and make sure the control level is the first level.
We will now show 4 ways of constructing a DESeqDataSet, depending on what pipeline was used upstream of DESeq2 to generated counts or estimated counts:
Our recommended pipeline for DESeq2 is to use fast transcript abundance quantifiers upstream of DESeq2, and then to create gene-level count matrices for use with DESeq2 by importing the quantification data using tximport (Soneson, Love, and Robinson 2015). This workflow allows users to import transcript abundance estimates from a variety of external software, including the following methods:
Some advantages of using the above methods for transcript abundance estimation are: (i) this approach corrects for potential changes in gene length across samples (e.g. from differential isoform usage) (Trapnell et al. 2013), (ii) some of these methods (Salmon, Sailfish, kallisto) are substantially faster and require less memory and disk usage compared to alignment-based methods that require creation and storage of BAM files, and (iii) it is possible to avoid discarding those fragments that can align to multiple genes with homologous sequence, thus increasing sensitivity (Robert and Watson 2015).
Full details on the motivation and methods for importing transcript level abundance and count estimates, summarizing to gene-level count matrices and producing an offset which corrects for potential changes in average transcript length across samples are described in (Soneson, Love, and Robinson 2015). Note that the tximport-to-DESeq2 approach uses estimated gene counts from the transcript abundance quantifiers, but not normalized counts.
A tutorial on how to use the Salmon software for quantifying transcript abundance can be found here. We recommend using the --gcBias
flag which estimates a correction factor for systematic biases commonly present in RNA-seq data (Love, Hogenesch, and Irizarry 2016; Patro et al. 2017), unless you are certain that your data do not contain such bias.
Here, we demonstrate how to import transcript abundances and construct a gene-level DESeqDataSet object from Salmon quant.sf
files, which are stored in the tximportData package. You do not need the tximportData
package for your analysis, it is only used here for demonstration.
Note that, instead of locating dir
using system.file, a user would typically just provide a path, e.g. /path/to/quant/files
. For a typical use, the condition
information should already be present as a column of the sample table samples
, while here we construct artificial condition labels for demonstration.
library("tximport")
library("readr")
library("tximportData")
dir <- system.file("extdata", package="tximportData")
samples <- read.table(file.path(dir,"samples.txt"), header=TRUE)
samples$condition <- factor(rep(c("A","B"),each=3))
rownames(samples) <- samples$run
samples[,c("pop","center","run","condition")]
## pop center run condition
## ERR188297 TSI UNIGE ERR188297 A
## ERR188088 TSI UNIGE ERR188088 A
## ERR188329 TSI UNIGE ERR188329 A
## ERR188288 TSI UNIGE ERR188288 B
## ERR188021 TSI UNIGE ERR188021 B
## ERR188356 TSI UNIGE ERR188356 B
Next we specify the path to the files using the appropriate columns of samples
, and we read in a table that links transcripts to genes for this dataset.
files <- file.path(dir,"salmon", samples$run, "quant.sf.gz")
names(files) <- samples$run
tx2gene <- read_csv(file.path(dir, "tx2gene.gencode.v27.csv"))
We import the necessary quantification data for DESeq2 using the tximport function. For further details on use of tximport, including the construction of the tx2gene
table for linking transcripts to genes in your dataset, please refer to the tximport package vignette.
Finally, we can construct a DESeqDataSet from the txi
object and sample information in samples
.
The ddsTxi
object here can then be used as dds
in the following analysis steps.
Another Bioconductor package, tximeta (Love et al. 2020), extends tximport, offering the same functionality, plus the additional benefit of automatic addition of annotation metadata for commonly used transcriptomes (GENCODE, Ensembl, RefSeq for human and mouse). See the tximeta package vignette for more details. tximeta produces a SummarizedExperiment that can be loaded easily into DESeq2 using the DESeqDataSet
function, with an example in the tximeta package vignette, and below:
The ddsTxi
object here can then be used as dds
in the following analysis steps. If tximeta recognized the reference transcriptome as one of those with a pre-computed hashed checksum, the rowRanges
of the dds
object will be pre-populated. Again, see the tximeta vignette for full details.
Alternatively, the function DESeqDataSetFromMatrix can be used if you already have a matrix of read counts prepared from another source. Another method for quickly producing count matrices from alignment files is the featureCounts function (Liao, Smyth, and Shi 2013) in the Rsubread package. To use DESeqDataSetFromMatrix, the user should provide the counts matrix, the information about the samples (the columns of the count matrix) as a DataFrame or data.frame, and the design formula.
To demonstrate the use of DESeqDataSetFromMatrix, we will read in count data from the pasilla package. We read in a count matrix, which we will name cts
, and the sample information table, which we will name coldata
. Further below we describe how to extract these objects from, e.g. featureCounts output.
library("pasilla")
pasCts <- system.file("extdata",
"pasilla_gene_counts.tsv",
package="pasilla", mustWork=TRUE)
pasAnno <- system.file("extdata",
"pasilla_sample_annotation.csv",
package="pasilla", mustWork=TRUE)
cts <- as.matrix(read.csv(pasCts,sep="\t",row.names="gene_id"))
coldata <- read.csv(pasAnno, row.names=1)
coldata <- coldata[,c("condition","type")]
coldata$condition <- factor(coldata$condition)
coldata$type <- factor(coldata$type)
We examine the count matrix and column data to see if they are consistent in terms of sample order.
## untreated1 untreated2 untreated3 untreated4 treated1 treated2
## FBgn0000003 0 0 0 0 0 0
## FBgn0000008 92 161 76 70 140 88
## treated3
## FBgn0000003 1
## FBgn0000008 70
## condition type
## treated1fb treated single-read
## treated2fb treated paired-end
## treated3fb treated paired-end
## untreated1fb untreated single-read
## untreated2fb untreated single-read
## untreated3fb untreated paired-end
## untreated4fb untreated paired-end
Note that these are not in the same order with respect to samples!
It is absolutely critical that the columns of the count matrix and the rows of the column data (information about samples) are in the same order. DESeq2 will not make guesses as to which column of the count matrix belongs to which row of the column data, these must be provided to DESeq2 already in consistent order.
As they are not in the correct order as given, we need to re-arrange one or the other so that they are consistent in terms of sample order (if we do not, later functions would produce an error). We additionally need to chop off the "fb"
of the row names of coldata
, so the naming is consistent.
## [1] TRUE
## [1] FALSE
## [1] TRUE
If you have used the featureCounts function (Liao, Smyth, and Shi 2013) in the Rsubread package, the matrix of read counts can be directly provided from the "counts"
element in the list output. The count matrix and column data can typically be read into R from flat files using base R functions such as read.csv or read.delim. For htseq-count files, see the dedicated input function below.
With the count matrix, cts
, and the sample information, coldata
, we can construct a DESeqDataSet:
library("DESeq2")
dds <- DESeqDataSetFromMatrix(countData = cts,
colData = coldata,
design = ~ condition)
dds
## class: DESeqDataSet
## dim: 14599 7
## metadata(1): version
## assays(1): counts
## rownames(14599): FBgn0000003 FBgn0000008 ... FBgn0261574 FBgn0261575
## rowData names(0):
## colnames(7): treated1 treated2 ... untreated3 untreated4
## colData names(2): condition type
If you have additional feature data, it can be added to the DESeqDataSet by adding to the metadata columns of a newly constructed object. (Here we add redundant data just for demonstration, as the gene names are already the rownames of the dds
.)
featureData <- data.frame(gene=rownames(cts))
mcols(dds) <- DataFrame(mcols(dds), featureData)
mcols(dds)
## DataFrame with 14599 rows and 1 column
## gene
## <character>
## FBgn0000003 FBgn0000003
## FBgn0000008 FBgn0000008
## FBgn0000014 FBgn0000014
## FBgn0000015 FBgn0000015
## FBgn0000017 FBgn0000017
## ... ...
## FBgn0261571 FBgn0261571
## FBgn0261572 FBgn0261572
## FBgn0261573 FBgn0261573
## FBgn0261574 FBgn0261574
## FBgn0261575 FBgn0261575
You can use the function DESeqDataSetFromHTSeqCount if you have used htseq-count from the HTSeq python package (Anders, Pyl, and Huber 2014). For an example of using the python scripts, see the pasilla data package. First you will want to specify a variable which points to the directory in which the htseq-count output files are located.
However, for demonstration purposes only, the following line of code points to the directory for the demo htseq-count output files packages for the pasilla package.
We specify which files to read in using list.files, and select those files which contain the string "treated"
using grep. The sub function is used to chop up the sample filename to obtain the condition status, or you might alternatively read in a phenotypic table using read.table.
sampleFiles <- grep("treated",list.files(directory),value=TRUE)
sampleCondition <- sub("(.*treated).*","\\1",sampleFiles)
sampleTable <- data.frame(sampleName = sampleFiles,
fileName = sampleFiles,
condition = sampleCondition)
sampleTable$condition <- factor(sampleTable$condition)
Then we build the DESeqDataSet using the following function:
library("DESeq2")
ddsHTSeq <- DESeqDataSetFromHTSeqCount(sampleTable = sampleTable,
directory = directory,
design= ~ condition)
ddsHTSeq
## class: DESeqDataSet
## dim: 70463 7
## metadata(1): version
## assays(1): counts
## rownames(70463): FBgn0000003:001 FBgn0000008:001 ... FBgn0261575:001
## FBgn0261575:002
## rowData names(0):
## colnames(7): treated1fb.txt treated2fb.txt ... untreated3fb.txt
## untreated4fb.txt
## colData names(1): condition
If one has already created or obtained a SummarizedExperiment, it can be easily input into DESeq2 as follows. First we load the package containing the airway
dataset.
The constructor function below shows the generation of a DESeqDataSet from a RangedSummarizedExperiment se
.
## class: DESeqDataSet
## dim: 63677 8
## metadata(2): '' version
## assays(1): counts
## rownames(63677): ENSG00000000003 ENSG00000000005 ... ENSG00000273492
## ENSG00000273493
## rowData names(10): gene_id gene_name ... seq_coord_system symbol
## colnames(8): SRR1039508 SRR1039509 ... SRR1039520 SRR1039521
## colData names(9): SampleName cell ... Sample BioSample
While it is not necessary to pre-filter low count genes before running the DESeq2 functions, there are two reasons which make pre-filtering useful: by removing rows in which there are very few reads, we reduce the memory size of the dds
data object, and we increase the speed of count modeling within DESeq2. It can also improve visualizations, as features with no information for differential expression are not plotted in dispersion plots or MA-plots.
Here we perform pre-filtering to keep only rows that have a count of at least 10 for a minimal number of samples. The count of 10 is a reasonable choice for bulk RNA-seq. A recommendation for the minimal number of samples is to specify the smallest group size, e.g. here there are 3 treated samples. If there are not discrete groups, one can use the minimal number of samples where non-zero counts would be considered interesting. One can also omit this step entirely and just rely on the independent filtering procedures available in results()
, either IHW or genefilter. See independent filtering section.
By default, R will choose a reference level for factors based on alphabetical order. Then, if you never tell the DESeq2 functions which level you want to compare against (e.g. which level represents the control group), the comparisons will be based on the alphabetical order of the levels. There are two solutions: you can either explicitly tell results which comparison to make using the contrast
argument (this will be shown later), or you can explicitly set the factors levels. In order to see the change of reference levels reflected in the results names, you need to either run DESeq
or nbinomWaldTest
/nbinomLRT
after the re-leveling operation. Setting the factor levels can be done in two ways, either using factor:
…or using relevel, just specifying the reference level:
If you need to subset the columns of a DESeqDataSet, i.e., when removing certain samples from the analysis, it is possible that all the samples for one or more levels of a variable in the design formula would be removed. In this case, the droplevels function can be used to remove those levels which do not have samples in the current DESeqDataSet:
DESeq2 provides a function collapseReplicates which can assist in combining the counts from technical replicates into single columns of the count matrix. The term technical replicate implies multiple sequencing runs of the same library. You should not collapse biological replicates using this function. See the manual page for an example of the use of collapseReplicates.
We continue with the pasilla data constructed from the count matrix method above. This data set is from an experiment on Drosophila melanogaster cell cultures and investigated the effect of RNAi knock-down of the splicing factor pasilla (Brooks et al. 2011). The detailed transcript of the production of the pasilla data is provided in the vignette of the data package pasilla.
The standard differential expression analysis steps are wrapped into a single function, DESeq. The estimation steps performed by this function are described below, in the manual page for ?DESeq
and in the Methods section of the DESeq2 publication (Love, Huber, and Anders 2014).
Results tables are generated using the function results, which extracts a results table with log2 fold changes, p values and adjusted p values. With no additional arguments to results, the log2 fold change and Wald test p value will be for the last variable in the design formula, and if this is a factor, the comparison will be the last level of this variable over the reference level (see previous note on factor levels). However, the order of the variables of the design do not matter so long as the user specifies the comparison to build a results table for, using the name
or contrast
arguments of results.
Details about the comparison are printed to the console, directly above the results table. The text, condition treated vs untreated
, tells you that the estimates are of the logarithmic fold change log2(treated/untreated).
## log2 fold change (MLE): condition treated vs untreated
## Wald test p-value: condition treated vs untreated
## DataFrame with 8148 rows and 6 columns
## baseMean log2FoldChange lfcSE stat pvalue padj
## <numeric> <numeric> <numeric> <numeric> <numeric> <numeric>
## FBgn0000008 95.28865 0.00399148 0.225010 0.0177391 0.9858470 0.996699
## FBgn0000017 4359.09632 -0.23842494 0.127094 -1.8759764 0.0606585 0.289604
## FBgn0000018 419.06811 -0.10185506 0.146568 -0.6949338 0.4870968 0.822681
## FBgn0000024 6.41105 0.21429657 0.691557 0.3098756 0.7566555 0.939146
## FBgn0000032 990.79225 -0.08896298 0.146253 -0.6082822 0.5430003 0.848881
## ... ... ... ... ... ... ...
## FBgn0261564 1160.028 -0.0857255 0.108354 -0.7911643 0.4288481 0.789246
## FBgn0261565 620.388 -0.2943294 0.140496 -2.0949303 0.0361772 0.206423
## FBgn0261570 3212.969 0.2971841 0.126742 2.3447877 0.0190379 0.133380
## FBgn0261573 2243.936 0.0146611 0.111365 0.1316493 0.8952617 0.977565
## FBgn0261574 4863.807 0.0179729 0.194137 0.0925784 0.9262385 0.986726
Note that we could have specified the coefficient or contrast we want to build a results table for, using either of the following equivalent commands:
res <- results(dds, name="condition_treated_vs_untreated")
res <- results(dds, contrast=c("condition","treated","untreated"))
One exception to the equivalence of these two commands, is that, using contrast
will additionally set to 0 the estimated LFC in a comparison of two groups, where all of the counts in the two groups are equal to 0 (while other groups have positive counts). As this may be a desired feature to have the LFC in these cases set to 0, one can use contrast
to build these results tables. More information about extracting specific coefficients from a fitted DESeqDataSet object can be found in the help page ?results
. The use of the contrast
argument is also further discussed below.
Shrinkage of effect size (LFC estimates) is useful for visualization and ranking of genes. To shrink the LFC, we pass the dds
object to the function lfcShrink
. Below we specify to use the apeglm method for effect size shrinkage (Zhu, Ibrahim, and Love 2018), which improves on the previous estimator.
We provide the dds
object and the name or number of the coefficient we want to shrink, where the number refers to the order of the coefficient as it appears in resultsNames(dds)
.
## [1] "Intercept" "condition_treated_vs_untreated"
## log2 fold change (MAP): condition treated vs untreated
## Wald test p-value: condition treated vs untreated
## DataFrame with 8148 rows and 5 columns
## baseMean log2FoldChange lfcSE pvalue padj
## <numeric> <numeric> <numeric> <numeric> <numeric>
## FBgn0000008 95.28865 0.00195376 0.152654 0.9858470 0.996699
## FBgn0000017 4359.09632 -0.18810628 0.120870 0.0606585 0.289604
## FBgn0000018 419.06811 -0.06893831 0.122805 0.4870968 0.822681
## FBgn0000024 6.41105 0.01786546 0.199499 0.7566555 0.939146
## FBgn0000032 990.79225 -0.06001511 0.121962 0.5430003 0.848881
## ... ... ... ... ... ...
## FBgn0261564 1160.028 -0.0669829 0.0976567 0.4288481 0.789246
## FBgn0261565 620.388 -0.2284564 0.1362122 0.0361772 0.206423
## FBgn0261570 3212.969 0.2395981 0.1237304 0.0190379 0.133380
## FBgn0261573 2243.936 0.0115395 0.0981689 0.8952617 0.977565
## FBgn0261574 4863.807 0.0101618 0.1417667 0.9262385 0.986726
Shrinkage estimation is discussed more in a later section.
The above steps should take less than 30 seconds for most analyses. For experiments with complex designs and many samples (e.g. dozens of coefficients, ~100s of samples), one may want to have faster computation than provided by the default run of DESeq
. We have two recommendations:
By using the argument fitType="glmGamPoi"
, one can leverage the faster NB GLM engine written by Constantin Ahlmann-Eltze. Note that glmGamPoi’s interface in DESeq2 requires use of test="LRT"
and specification of a reduced
design.
One can take advantage of parallelized computation. Parallelizing DESeq
, results
, and lfcShrink
can be easily accomplished by loading the BiocParallel package, and then setting the following arguments: parallel=TRUE
and BPPARAM=MulticoreParam(4)
, for example, splitting the job over 4 cores. However, some words of advice on parallelization: first, it is recommend to filter genes where all samples have low counts, to avoid sending data unnecessarily to child processes, when those genes have low power and will be independently filtered anyway; secondly, there is often diminishing returns for adding more cores due to overhead of sending data to child processes, therefore I recommend first starting with small number of additional cores. Note that obtaining results
for coefficients or contrasts listed in resultsNames(dds)
is fast and will not need parallelization. As an alternative to BPPARAM
, one can register
cores at the beginning of an analysis, and then just specify parallel=TRUE
to the functions when called.
We can order our results table by the smallest p value:
We can summarize some basic tallies using the summary function.
##
## out of 8148 with nonzero total read count
## adjusted p-value < 0.1
## LFC > 0 (up) : 533, 6.5%
## LFC < 0 (down) : 536, 6.6%
## outliers [1] : 0, 0%
## low counts [2] : 0, 0%
## (mean count < 5)
## [1] see 'cooksCutoff' argument of ?results
## [2] see 'independentFiltering' argument of ?results
How many adjusted p-values were less than 0.1?
## [1] 1069
The results function contains a number of arguments to customize the results table which is generated. You can read about these arguments by looking up ?results
. Note that the results function automatically performs independent filtering based on the mean of normalized counts for each gene, optimizing the number of genes which will have an adjusted p value below a given FDR cutoff, alpha
. Independent filtering is further discussed below. By default the argument alpha
is set to \(0.1\). If the adjusted p value cutoff will be a value other than \(0.1\), alpha
should be set to that value:
##
## out of 8148 with nonzero total read count
## adjusted p-value < 0.05
## LFC > 0 (up) : 416, 5.1%
## LFC < 0 (down) : 437, 5.4%
## outliers [1] : 0, 0%
## low counts [2] : 0, 0%
## (mean count < 5)
## [1] see 'cooksCutoff' argument of ?results
## [2] see 'independentFiltering' argument of ?results
## [1] 853
A generalization of the idea of p value filtering is to weight hypotheses to optimize power. A Bioconductor package, IHW, is available that implements the method of Independent Hypothesis Weighting (Ignatiadis et al. 2016). Here we show the use of IHW for p value adjustment of DESeq2 results. For more details, please see the vignette of the IHW package. The IHW result object is stored in the metadata.
Note: If the results of independent hypothesis weighting are used in published research, please cite:
Ignatiadis, N., Klaus, B., Zaugg, J.B., Huber, W. (2016) Data-driven hypothesis weighting increases detection power in genome-scale multiple testing. Nature Methods, 13:7. 10.1038/nmeth.3885
# (unevaluated code chunk)
library("IHW")
resIHW <- results(dds, filterFun=ihw)
summary(resIHW)
sum(resIHW$padj < 0.1, na.rm=TRUE)
metadata(resIHW)$ihwResult
For advanced users, note that all the values calculated by the DESeq2 package are stored in the DESeqDataSet object or the DESeqResults object, and access to these values is discussed below.
In DESeq2, the function plotMA shows the log2 fold changes attributable to a given variable over the mean of normalized counts for all the samples in the DESeqDataSet. Points will be colored blue if the adjusted p value is less than 0.1. Points which fall out of the window are plotted as open triangles pointing either up or down.
It is more useful to visualize the MA-plot for the shrunken log2 fold changes, which remove the noise associated with log2 fold changes from low count genes without requiring arbitrary filtering thresholds.
After calling plotMA, one can use the function identify to interactively detect the row number of individual genes by clicking on the plot. One can then recover the gene identifiers by saving the resulting indices:
The moderated log fold changes proposed by Love, Huber, and Anders (2014) use a normal prior distribution, centered on zero and with a scale that is fit to the data. The shrunken log fold changes are useful for ranking and visualization, without the need for arbitrary filters on low count genes. The normal prior can sometimes produce too strong of shrinkage for certain datasets. In DESeq2 version 1.18, we include two additional adaptive shrinkage estimators, available via the type
argument of lfcShrink
. For more details, see ?lfcShrink
The options for type
are:
apeglm
is the adaptive t prior shrinkage estimator from the apeglm package (Zhu, Ibrahim, and Love 2018). As of version 1.28.0, it is the default estimator.ashr
is the adaptive shrinkage estimator from the ashr package (Stephens 2016). Here DESeq2 uses the ashr option to fit a mixture of Normal distributions to form the prior, with method="shrinkage"
.normal
is the the original DESeq2 shrinkage estimator, an adaptive Normal distribution as prior.If the shrinkage estimator apeglm
is used in published research, please cite:
Zhu, A., Ibrahim, J.G., Love, M.I. (2018) Heavy-tailed prior distributions for sequence count data: removing the noise and preserving large differences. Bioinformatics. 10.1093/bioinformatics/bty895
If the shrinkage estimator ashr
is used in published research, please cite:
Stephens, M. (2016) False discovery rates: a new deal. Biostatistics, 18:2. 10.1093/biostatistics/kxw041
In the LFC shrinkage code above, we specified coef="condition_treated_vs_untreated"
. We can also just specify the coefficient by the order that it appears in resultsNames(dds)
, in this case coef=2
. For more details explaining how the shrinkage estimators differ, and what kinds of designs, contrasts and output is provided by each, see the extended section on shrinkage estimators.
## [1] "Intercept" "condition_treated_vs_untreated"
# because we are interested in treated vs untreated, we set 'coef=2'
resNorm <- lfcShrink(dds, coef=2, type="normal")
resAsh <- lfcShrink(dds, coef=2, type="ashr")
par(mfrow=c(1,3), mar=c(4,4,2,1))
xlim <- c(1,1e5); ylim <- c(-3,3)
plotMA(resLFC, xlim=xlim, ylim=ylim, main="apeglm")
plotMA(resNorm, xlim=xlim, ylim=ylim, main="normal")
plotMA(resAsh, xlim=xlim, ylim=ylim, main="ashr")
Note: We have sped up the apeglm
method so it takes roughly about the same amount of time as normal
, e.g. ~5 seconds for the pasilla
dataset of ~10,000 genes and 7 samples. If fast shrinkage estimation of LFC is needed, but the posterior standard deviation is not needed, setting apeMethod="nbinomC"
will produce a ~10x speedup, but the lfcSE
column will be returned with NA
. A variant of this fast method, apeMethod="nbinomC*"
includes random starts.
Note: If there is unwanted variation present in the data (e.g. batch effects) it is always recommend to correct for this, which can be accommodated in DESeq2 by including in the design any known batch variables or by using functions/packages such as svaseq
in sva (Leek 2014) or the RUV
functions in RUVSeq (Risso et al. 2014) to estimate variables that capture the unwanted variation. In addition, the ashr developers have a specific method for accounting for unwanted variation in combination with ashr (Gerard and Stephens 2017).
It can also be useful to examine the counts of reads for a single gene across the groups. A simple function for making this plot is plotCounts, which normalizes counts by the estimated size factors (or normalization factors if these were used) and adds a pseudocount of 1/2 to allow for log scale plotting. The counts are grouped by the variables in intgroup
, where more than one variable can be specified. Here we specify the gene which had the smallest p value from the results table created above. You can select the gene to plot by rowname or by numeric index.
For customized plotting, an argument returnData
specifies that the function should only return a data.frame for plotting with ggplot.
Information about which variables and tests were used can be found by calling the function mcols on the results object.
## [1] "mean of normalized counts for all samples"
## [2] "log2 fold change (MLE): condition treated vs untreated"
## [3] "standard error: condition treated vs untreated"
## [4] "Wald statistic: condition treated vs untreated"
## [5] "Wald test p-value: condition treated vs untreated"
## [6] "BH adjusted p-values"
For a particular gene, a log2 fold change of -1 for condition treated vs untreated
means that the treatment induces a multiplicative change in observed gene expression level of \(2^{-1} = 0.5\) compared to the untreated condition. If the variable of interest is continuous-valued, then the reported log2 fold change is per unit of change of that variable.
Note on p-values set to NA: some values in the results table can be set to NA
for one of the following reasons:
baseMean
column will be zero, and the log2 fold change estimates, p value and adjusted p value will all be set to NA
.NA
. These outlier counts are detected by Cook’s distance. Customization of this outlier filtering and description of functionality for replacement of outlier counts and refitting is described belowNA
. Description and customization of independent filtering is described belowregionReport An HTML and PDF summary of the results with plots can also be generated using the regionReport package. The DESeq2Report function should be run on a DESeqDataSet that has been processed by the DESeq function. For more details see the manual page for DESeq2Report and an example vignette in the regionReport package.
Glimma Interactive visualization of DESeq2 output, including MA-plots (also called MD-plots) can be generated using the Glimma package. See the manual page for glMDPlot.DESeqResults.
pcaExplorer Interactive visualization of DESeq2 output, including PCA plots, boxplots of counts and other useful summaries can be generated using the pcaExplorer package. See the Launching the application section of the package vignette.
iSEE Provides functions for creating an interactive Shiny-based graphical user interface for exploring data stored in SummarizedExperiment objects, including row- and column-level metadata. Particular attention is given to single-cell data in a SingleCellExperiment object with visualization of dimensionality reduction results. iSEE is on Bioconductor. An example wrapper function for converting a DESeqDataSet to a SingleCellExperiment object for use with iSEE can be found at the following gist, written by Federico Marini:
The iSEEde package provides additional panels that facilitate the interactive visualisation of differential expression results in iSEE applications.
DEvis DEvis is a powerful, integrated solution for the analysis of differential expression data. This package includes an array of tools for manipulating and aggregating data, as well as a wide range of customizable visualizations, and project management functionality that simplify RNA-Seq analysis and provide a variety of ways of exploring and analyzing data. DEvis can be found on CRAN and GitHub.
A plain-text file of the results can be exported using the base R functions write.csv or write.delim. We suggest using a descriptive file name indicating the variable and levels which were tested.
Exporting only the results which pass an adjusted p value threshold can be accomplished with the subset function, followed by the write.csv function.
## log2 fold change (MLE): condition treated vs untreated
## Wald test p-value: condition treated vs untreated
## DataFrame with 1069 rows and 6 columns
## baseMean log2FoldChange lfcSE stat pvalue
## <numeric> <numeric> <numeric> <numeric> <numeric>
## FBgn0039155 730.992 -4.61695 0.1667827 -27.6824 1.13645e-168
## FBgn0025111 1504.272 2.90205 0.1261989 22.9958 5.13139e-117
## FBgn0029167 3709.741 -2.19491 0.0957474 -22.9240 2.68013e-116
## FBgn0003360 4344.597 -3.17683 0.1416166 -22.4326 1.89322e-111
## FBgn0035085 638.757 -2.55819 0.1359972 -18.8106 6.18406e-79
## ... ... ... ... ... ...
## FBgn0003890 1644.2002 -0.495185 0.199269 -2.48501 0.0129549
## FBgn0010053 178.3259 -0.516894 0.208034 -2.48466 0.0129675
## FBgn0034997 5125.8312 0.250348 0.100801 2.48360 0.0130063
## FBgn0004359 84.1178 0.646359 0.260308 2.48306 0.0130260
## FBgn0027604 283.2465 -0.578675 0.233077 -2.48277 0.0130366
## padj
## <numeric>
## FBgn0039155 9.25983e-165
## FBgn0025111 2.09053e-113
## FBgn0029167 7.27922e-113
## FBgn0003360 3.85649e-108
## FBgn0035085 1.00775e-75
## ... ...
## FBgn0003890 0.0991143
## FBgn0010053 0.0991176
## FBgn0034997 0.0993210
## FBgn0004359 0.0993659
## FBgn0027604 0.0993659
Experiments with more than one factor influencing the counts can be analyzed using design formulas that include the additional variables. In fact, DESeq2 can analyze any possible experimental design that can be expressed with fixed effects terms (multiple factors, designs with interactions, designs with continuous variables, splines, and so on are all possible).
By adding variables to the design, one can control for additional variation in the counts. For example, if the condition samples are balanced across experimental batches, by including the batch
factor to the design, one can increase the sensitivity for finding differences due to condition
. There are multiple ways to analyze experiments when the additional variables are of interest and not just controlling factors (see section on interactions).
Experiments with many samples: in experiments with many samples (e.g. 50, 100, etc.) it is highly likely that there will be technical variation affecting the observed counts. Failing to model this additional technical variation will lead to spurious results. Many methods exist that can be used to model technical variation, which can be easily included in the DESeq2 design to control for technical variation while estimating effects of interest. See the RNA-seq workflow for examples of using RUV or SVA in combination with DESeq2. For more details on why it is important to control for technical variation in large sample experiments, see the following thread, also archived here by Frederik Ziebell.
The data in the pasilla package have a condition of interest (the column condition
), as well as information on the type of sequencing which was performed (the column type
), as we can see below:
## DataFrame with 7 rows and 3 columns
## condition type sizeFactor
## <factor> <factor> <numeric>
## treated1 treated single-read 1.629707
## treated2 treated paired-end 0.761162
## treated3 treated paired-end 0.830312
## untreated1 untreated single-read 1.143904
## untreated2 untreated single-read 1.791281
## untreated3 untreated paired-end 0.645994
## untreated4 untreated paired-end 0.750728
We create a copy of the DESeqDataSet, so that we can rerun the analysis using a multi-factor design.
We change the levels of type
so it only contains letters (numbers, underscore and period are also allowed in design factor levels). Be careful when changing level names to use the same order as the current levels.
## [1] "paired-end" "single-read"
## [1] "paired" "single"
We can account for the different types of sequencing, and get a clearer picture of the differences attributable to the treatment. As condition
is the variable of interest, we put it at the end of the formula. Thus the results function will by default pull the condition
results unless contrast
or name
arguments are specified.
Then we can re-run DESeq:
Again, we access the results using the results function.
## log2 fold change (MLE): condition treated vs untreated
## Wald test p-value: condition treated vs untreated
## DataFrame with 6 rows and 6 columns
## baseMean log2FoldChange lfcSE stat pvalue padj
## <numeric> <numeric> <numeric> <numeric> <numeric> <numeric>
## FBgn0000008 95.28865 -0.0390130 0.218997 -0.178144 0.8586100 0.947833
## FBgn0000017 4359.09632 -0.2548984 0.113535 -2.245099 0.0247617 0.131475
## FBgn0000018 419.06811 -0.0625571 0.129956 -0.481372 0.6302523 0.852180
## FBgn0000024 6.41105 0.3097331 0.750231 0.412850 0.6797164 0.877741
## FBgn0000032 990.79225 -0.0465134 0.120215 -0.386918 0.6988171 0.886082
## FBgn0000037 14.11443 0.4541562 0.523436 0.867644 0.3855893 0.691941
It is also possible to retrieve the log2 fold changes, p values and adjusted p values of variables other than the last one in the design. While in this case, type
is not biologically interesting as it indicates differences across sequencing protocol, for other hypothetical designs, such as ~genotype + condition + genotype:condition
, we may actually be interested in the difference in baseline expression across genotype, which is not the last variable in the design.
In any case, the contrast
argument of the function results takes a character vector of length three: the name of the variable, the name of the factor level for the numerator of the log2 ratio, and the name of the factor level for the denominator. The contrast
argument can also take other forms, as described in the help page for results and below
## log2 fold change (MLE): type single vs paired
## Wald test p-value: type single vs paired
## DataFrame with 6 rows and 6 columns
## baseMean log2FoldChange lfcSE stat pvalue padj
## <numeric> <numeric> <numeric> <numeric> <numeric> <numeric>
## FBgn0000008 95.28865 -0.265123 0.217459 -1.219189 0.2227725 0.492729
## FBgn0000017 4359.09632 -0.103203 0.113399 -0.910090 0.3627748 0.642817
## FBgn0000018 419.06811 0.225857 0.128864 1.752669 0.0796589 0.271610
## FBgn0000024 6.41105 0.302083 0.745703 0.405099 0.6854049 NA
## FBgn0000032 990.79225 0.233891 0.119646 1.954855 0.0506002 0.206081
## FBgn0000037 14.11443 -0.053260 0.521939 -0.102043 0.9187228 0.969866
If the variable is continuous or an interaction term (see section on interactions) then the results can be extracted using the name
argument to results, where the name is one of elements returned by resultsNames(dds)
.
In order to test for differential expression, we operate on raw counts and use discrete distributions as described in the previous section on differential expression. However for other downstream analyses – e.g. for visualization or clustering – it might be useful to work with transformed versions of the count data.
Maybe the most obvious choice of transformation is the logarithm. Since count values for a gene can be zero in some conditions (and non-zero in others), some advocate the use of pseudocounts, i.e. transformations of the form:
\[ y = \log_2(n + n_0) \]
where n represents the count values and \(n_0\) is a positive constant.
In this section, we discuss two alternative approaches that offer more theoretical justification and a rational way of choosing parameters equivalent to \(n_0\) above. One makes use of the concept of variance stabilizing transformations (VST) (Tibshirani 1988; Huber et al. 2003; Anders and Huber 2010), and the other is the regularized logarithm or rlog, which incorporates a prior on the sample differences (Love, Huber, and Anders 2014). Both transformations produce transformed data on the log2 scale which has been normalized with respect to library size or other normalization factors.
The point of these two transformations, the VST and the rlog, is to remove the dependence of the variance on the mean, particularly the high variance of the logarithm of count data when the mean is low. Both VST and rlog use the experiment-wide trend of variance over mean, in order to transform the data to remove the experiment-wide trend. Note that we do not require or desire that all the genes have exactly the same variance after transformation. Indeed, in a figure below, you will see that after the transformations the genes with the same mean do not have exactly the same standard deviations, but that the experiment-wide trend has flattened. It is those genes with row variance above the trend which will allow us to cluster samples into interesting groups.
Note on running time: if you have many samples (e.g. 100s), the rlog function might take too long, and so the vst function will be a faster choice. The rlog and VST have similar properties, but the rlog requires fitting a shrinkage term for each sample and each gene which takes time. See the DESeq2 paper for more discussion on the differences (Love, Huber, and Anders 2014).
The two functions, vst and rlog have an argument blind
, for whether the transformation should be blind to the sample information specified by the design formula. When blind
equals TRUE
(the default), the functions will re-estimate the dispersions using only an intercept. This setting should be used in order to compare samples in a manner wholly unbiased by the information about experimental groups, for example to perform sample QA (quality assurance) as demonstrated below.
However, blind dispersion estimation is not the appropriate choice if one expects that many or the majority of genes (rows) will have large differences in counts which are explainable by the experimental design, and one wishes to transform the data for downstream analysis. In this case, using blind dispersion estimation will lead to large estimates of dispersion, as it attributes differences due to experimental design as unwanted noise, and will result in overly shrinking the transformed values towards each other. By setting blind
to FALSE
, the dispersions already estimated will be used to perform transformations, or if not present, they will be estimated using the current design formula. Note that only the fitted dispersion estimates from mean-dispersion trend line are used in the transformation (the global dependence of dispersion on mean for the entire experiment). So setting blind
to FALSE
is still for the most part not using the information about which samples were in which experimental group in applying the transformation.
These transformation functions return an object of class DESeqTransform which is a subclass of RangedSummarizedExperiment. For ~20 samples, running on a newly created DESeqDataSet
, rlog may take 30 seconds, while vst takes less than 1 second. The running times are shorter when using blind=FALSE
and if the function DESeq has already been run, because then it is not necessary to re-estimate the dispersion values. The assay function is used to extract the matrix of normalized values.
## treated1 treated2 treated3 untreated1 untreated2 untreated3
## FBgn0000008 7.777746 7.984491 7.765448 7.735026 7.807720 7.997378
## FBgn0000017 11.954934 12.035700 12.028610 12.049838 12.295662 12.471723
## FBgn0000018 9.219162 9.089390 9.028791 9.374577 9.173538 9.052143
## untreated4
## FBgn0000008 7.832458
## FBgn0000017 12.089675
## FBgn0000018 9.142848
Above, we used a parametric fit for the dispersion. In this case, the closed-form expression for the variance stabilizing transformation is used by the vst function. If a local fit is used (option fitType="locfit"
to estimateDispersions) a numerical integration is used instead. The transformed data should be approximated variance stabilized and also includes correction for size factors or normalization factors. The transformed data is on the log2 scale for large counts.
The function rlog, stands for regularized log, transforming the original count data to the log2 scale by fitting a model with a term for each sample and a prior distribution on the coefficients which is estimated from the data. This is the same kind of shrinkage (sometimes referred to as regularization, or moderation) of log fold changes used by DESeq and nbinomWaldTest. The resulting data contains elements defined as:
\[ \log_2(q_{ij}) = \beta_{i0} + \beta_{ij} \]
where \(q_{ij}\) is a parameter proportional to the expected true concentration of fragments for gene i and sample j (see formula below), \(\beta_{i0}\) is an intercept which does not undergo shrinkage, and \(\beta_{ij}\) is the sample-specific effect which is shrunk toward zero based on the dispersion-mean trend over the entire dataset. The trend typically captures high dispersions for low counts, and therefore these genes exhibit higher shrinkage from the rlog.
Note that, as \(q_{ij}\) represents the part of the mean value \(\mu_{ij}\) after the size factor \(s_j\) has been divided out, it is clear that the rlog transformation inherently accounts for differences in sequencing depth. Without priors, this design matrix would lead to a non-unique solution, however the addition of a prior on non-intercept betas allows for a unique solution to be found.
The figure below plots the standard deviation of the transformed data, across samples, against the mean, using the shifted logarithm transformation, the regularized log transformation and the variance stabilizing transformation. The shifted logarithm has elevated standard deviation in the lower count range, and the regularized log to a lesser extent, while for the variance stabilized data the standard deviation is roughly constant along the whole dynamic range.
Note that the vertical axis in such plots is the square root of the variance over all samples, so including the variance due to the experimental conditions. While a flat curve of the square root of variance over the mean may seem like the goal of such transformations, this may be unreasonable in the case of datasets with many true differences due to the experimental conditions.