1 Using DESeq2 for Pairwise Differential Expression

1.1 Rationale

DESeq2’s default treatment of data relies on the assumption that genewise estimates of dispersion are largely unchanged across samples. While this assumption holds for a typical RNA-seq data, it can be violated if there are samples within the DESeqDataSet object for which there are meaningful signal changes across a majority of regions of interest.

The BRGenomics functions getDESeqDataSet() and getDESeqResults() are simple and flexible wrappers for making pairwise comparisons between individual samples, without relying on the assumption of globally-similar dispersion estimates. In particular, getDESeqResults() follows the logic that the presence of a dataset \(X\) within the DESeqDataSet object will not affect the comparison of datasets \(Y\) and \(Z\).

While the intuition above is appealing, users should note that if the globally-similar dispersions assumption does hold, then DESeq2’s default behavior should be more sensitive in its estimates of genewise dispersion. In this case, users can still take advantage of the convenience of the BRGenomics function getDESeqDataSet(), but they should subsequently call DESeq2::DESeq() and DESeq2::results() directly.

If the globally-similar dispersions assumption is violated, but something beyond simple pairwise comparisons is desired (e.g. group comparisons or additional model terms), we note that, with some prying, DESeq2 can be used without “blind dispersion estimation” (see the DESeq2 manual).

1.2 Formatting Data for DESeq2

Just like the functions that generate batch-normalized spike-in normalization factors, the DESeq-oriented functions require that the names of the input datasets end in "rep1", "rep2", etc.

Let’s first make a toy list of multiple datasets to compare:

library(BRGenomics)
data("PROseq")
ps_list <- lapply(1:6, function(i) PROseq[seq(i, length(PROseq), 6)])
names(ps_list) <- c("A_rep1", "A_rep2", 
                    "B_rep1", "B_rep2",
                    "C_rep1", "C_rep2")
ps_list[1:2]
## $A_rep1
## GRanges object with 7897 ranges and 1 metadata column:
##          seqnames    ranges strand |     score
##             <Rle> <IRanges>  <Rle> | <integer>
##      [1]     chr4      1295      + |         1
##      [2]     chr4     42595      + |         1
##      [3]     chr4     42622      + |         2
##      [4]     chr4     42718      + |         1
##      [5]     chr4     42789      + |         1
##      ...      ...       ...    ... .       ...
##   [7893]     chr4   1295277      - |         1
##   [7894]     chr4   1306500      - |         1
##   [7895]     chr4   1306889      - |         1
##   [7896]     chr4   1307122      - |         1
##   [7897]     chr4   1316537      - |         1
##   -------
##   seqinfo: 7 sequences from an unspecified genome
## 
## $A_rep2
## GRanges object with 7897 ranges and 1 metadata column:
##          seqnames    ranges strand |     score
##             <Rle> <IRanges>  <Rle> | <integer>
##      [1]     chr4     41428      + |         1
##      [2]     chr4     42596      + |         1
##      [3]     chr4     42652      + |         3
##      [4]     chr4     42733      + |         1
##      [5]     chr4     42794      + |         2
##      ...      ...       ...    ... .       ...
##   [7893]     chr4   1305972      - |         1
##   [7894]     chr4   1306514      - |         1
##   [7895]     chr4   1307032      - |         1
##   [7896]     chr4   1307126      - |         1
##   [7897]     chr4   1318960      - |         1
##   -------
##   seqinfo: 7 sequences from an unspecified genome
names(ps_list)
## [1] "A_rep1" "A_rep2" "B_rep1" "B_rep2" "C_rep1" "C_rep2"

As you can see, the names all end in “repX”, where X indicates the replicate. Replicates will be grouped by anything that follows “rep”. If the sample names do not conform to this standard, the sample_names argument can be used to rename the samples within the call to getDESeqDataSet().

data("txs_dm6_chr4")
dds <- getDESeqDataSet(ps_list, txs_dm6_chr4,
                       gene_names = txs_dm6_chr4$gene_id,
                       ncores = 1)
dds
## class: DESeqDataSet 
## dim: 111 6 
## metadata(1): version
## assays(1): counts
## rownames(111): FBgn0267363 FBgn0266617 ... FBgn0039924 FBgn0027101
## rowData names(2): tx_name gene_id
## colnames(6): A_rep1 A_rep2 ... C_rep1 C_rep2
## colData names(2): condition replicate

Notice that the dim attribute of the DESeqDataSet object is c(111, 6). There are 6 samples, but length(txs_dm6_chr4) is not 111. This is because we provided gene names to getDESeqDataSet(), which were non-unique. The feature being exploited here is for use with discontinuous gene regions, not for multiple overlapping transcript isoforms.


By default, getDESeqDataSet() will combine counts across all ranges belonging to a gene, but if they overlap, they will be counted twice. For addressing issues related to overlaps, see the reduceByGene() and intersectByGene() functions.


We could have added normalization factors, which DESeq2 calls “size factors”, in the call to getDESeqDataSet(), or we can supply them to getDESeqResults() below. (Supplying them twice will overwrite the previous size factors).

Important note on normalization factors: DESeq2 “size factors” are the inverse of BRGenomics normalization factors. So if you calculate normalization factors with NF <- getSpikeInNFs(...), set sizeFactors <- 1/NF.

1.3 Getting DESeq2 Results

Generating DESeq2 results is simple:

getDESeqResults(dds, contrast.numer = "B", contrast.denom = "A",
                quiet = TRUE, ncores = 1)
## log2 fold change (MLE): condition B vs A 
## Wald test p-value: condition B vs A 
## DataFrame with 111 rows and 6 columns
##                baseMean log2FoldChange     lfcSE       stat    pvalue      padj
##               <numeric>      <numeric> <numeric>  <numeric> <numeric> <numeric>
## FBgn0267363    0.252443     -1.4201507  4.997269 -0.2841853  0.776268  0.990374
## FBgn0266617    9.648855     -0.4357749  0.983164 -0.4432374  0.657594  0.990374
## FBgn0265633    2.291440      0.3467829  1.882149  0.1842484  0.853819  0.990374
## FBgn0264617   67.131764      0.0245325  0.451334  0.0543556  0.956652  0.990374
## FBgn0085432 4688.232636     -0.0469812  0.270431 -0.1737268  0.862080  0.990374
## ...                 ...            ...       ...        ...       ...       ...
## FBgn0039928     653.783    -0.00994936  0.262637 -0.0378825  0.969781  0.990374
## FBgn0052017     114.906     0.02350299  0.369082  0.0636796  0.949225  0.990374
## FBgn0002521     168.778     0.05777885  0.347249  0.1663903  0.867850  0.990374
## FBgn0039924     170.213    -0.20850730  0.409865 -0.5087219  0.610947  0.990374
## FBgn0027101     343.009    -0.21151750  0.290061 -0.7292177  0.465868  0.990374

We can also make multiple pairwise-comparisons by ignoring the contrast.numer and contrast.denom arguments, and instead using the comparisons argument. The resulting list of results is named according to the comparisons:

comparison_list <- list(c("B", "A"), 
                        c("C", "A"),
                        c("C", "B"))
dsres <- getDESeqResults(dds, comparisons = comparison_list,
                         quiet = TRUE, ncores = 1)
names(dsres)
## [1] "B_vs_A" "C_vs_A" "C_vs_B"
dsres$C_vs_B
## log2 fold change (MLE): condition C vs B 
## Wald test p-value: condition C vs B 
## DataFrame with 111 rows and 6 columns
##               baseMean log2FoldChange     lfcSE       stat    pvalue      padj
##              <numeric>      <numeric> <numeric>  <numeric> <numeric> <numeric>
## FBgn0267363    0.00000             NA        NA         NA        NA        NA
## FBgn0266617    9.38822      0.3938871  0.901764   0.436796  0.662259  0.998692
## FBgn0265633    2.28968     -0.3206785  1.718554  -0.186598  0.851976  0.998692
## FBgn0264617   65.30695     -0.0773192  0.415908  -0.185905  0.852520  0.998692
## FBgn0085432 4281.29059     -0.1902162  0.229704  -0.828094  0.407617  0.998692
## ...                ...            ...       ...        ...       ...       ...
## FBgn0039928    696.482      0.2150507  0.248408  0.8657159  0.386646  0.998692
## FBgn0052017    133.743      0.4160583  0.333948  1.2458762  0.212810  0.998692
## FBgn0002521    171.141      0.0159143  0.323995  0.0491188  0.960825  0.998692
## FBgn0039924    138.749     -0.3685073  0.327200 -1.1262461  0.260061  0.998692
## FBgn0027101    347.104      0.2699312  0.271542  0.9940675  0.320190  0.998692