How does DESeq2 "collapseReplicates()" function work on read counts data?

Question

Comparing read counts from an RNA-seq experiment for two select genes before and after using DESeq2's collapseReplicates() and plotCounts() functions yields interesting results:

Before collapseReplicates() and plotCounts():

Geneid	foo1.1	foo1.2	foo2.1	foo2.2	bar1.1	bar1.2	bar2.1	bar2.2	baz1.1	baz1.2	baz2.1	baz2.2	baz3.1	baz3.2
WASH7P	6	5	0	2	1	1	8	5	0	0	0	0	0	0
SOX3	1880	1861	1950	2055	1189	1181	2415	2482	3887	3810	1851	1738	3217	3406

After collapseReplicates() and plotCounts():

Geneid	foo1	foo2	bar1	bar2	baz1	baz2	baz3
WASH7P	9.877191	2.279384	3.478891	11.613875	0.500000	0.500000	0.500000
SOX3	3189.598	3563.717	3530.486	4187.011	6473.122	7991.460	5390.721

---

Note: In the above comparison, there are two (2) technical replicates (1.1 + 1.2, 2.1 + 2.2, etc.) for each biological replicate (foo1, foo2, etc.) for each of three (3) conditions (foo, bar, and baz).

---

Comparing the tables above, it appears as though DESeq2 is NOT taking the average or sum of columns being collapsed.

It is also curious - and mildly concerning - that some very low expression (i.e., 0.5 counts) is reported for genes in the matrix of collapsed replicates when, in the original count matrix, zero (0) reads were counted as 'mapped to that gene'.

So, how does collapseReplicates() "combine counts into single columns of the count matrix" as is described in the DESeq2 vignette?

---

Here is the code to collapse replicates and retrieve the number of read counts for a specific gene (e.g., WASH7P) in a dataframe to be used in a count plot, getting the "after" results shown above:

dds <- DESeqDataSetFromMatrix(countData = cts, colData = coldata, design = ~ condition)
ddsColl <- collapseReplicates(dds, dds$sample, dds$run)
keep <- rowSums(counts(ddsColl)) >= 10
dds <- ddsColl[keep,]
dds$condition <- relevel(dds$condition, ref = "baz")
dds <- DESeq(dds)
countsdf <- plotCounts(dds, gene="WASH7P", intgroup="condition", returnData=TRUE)

Answer 1

The DESeq2 function collapseReplicates sums the counts for the technical replicates.
Here is the code reference: github.com/mikelove/DESeq2/blob/master/R/helper.R#L186

OPs actual confusion was with the DESeq2 function plotCounts which by default normalizes count data and adds a pseudocount of 0.5 for plotting on log2 scale.