paired-end vs. single end for differential expression analysis

[nullalleles]

There are a variety of posts suggesting either single-end or paired-end mRNA-sequencing is better for differential expression analysis. I have noticed many people defaulting to paired-end reads, (even for RNA-seq, without assembly), but the reasons are not clear.

Is anyone aware of work that has been done to robustly address this question (perhaps with Illumina)?

It would be interesting to know the differences in costs, relative to alignable bases, as well as if there are any benefits of one or another solely for downstream differential expression analysis.

[Torst]

I think there are a few reasons people use PE for RNA-Seq:

1. Most sequencing centres only do PE 100bp so there is no choice

2. Illumina pricing PE doesn't cost that much more than SE (changing with MiSeq)

3. Many people want to do DGE, but may want to also denovo assemble, where PE is better

4. PE reads might be more uniquely alignable to the genome than a SE read. And for DGE, that is important. However PE does not give more statistical power than SE, as all that matters is counting tags, and a PE reads and a SE read both count as a single tag.

FYI - for bacteria, we try to use SE 50bp for RNA-Seq, as we don't have isoform/splice variant issues, and 50bp is enough to map uniquely.

[ymc]

PE is also better in finding fusion genes.

[vehuardo]

50 bp single end sequencing

Hi Torst,

I see this reply of yours is rather old, but I´d like to follow up on this: do you use 50 bp SE sequencing for bacterial RNA-seq?

I am doing some bacterial RNA-seq myself, ran a few samples on MiSeq 2x150. I will now do more samples on the HiSeq, but not really sure that 2x100 is worth the additional cost....

Best,

Vegard

I think there are a few reasons people use PE for RNA-Seq:

1. Most sequencing centres only do PE 100bp so there is no choice

2. Illumina pricing PE doesn't cost that much more than SE (changing with MiSeq)

3. Many people want to do DGE, but may want to also denovo assemble, where PE is better

4. PE reads might be more uniquely alignable to the genome than a SE read. And for DGE, that is important. However PE does not give more statistical power than SE, as all that matters is counting tags, and a PE reads and a SE read both count as a single tag.

FYI - for bacteria, we try to use SE 50bp for RNA-Seq, as we don't have isoform/splice variant issues, and 50bp is enough to map uniquely.

[james hadfield]

I'd recommend SE50bp for most DGE experiments as they are cheap and plentiful. More sample replicates is better than more reads to estimate within group variability better. We have done work on this but need ot pull our fingers out and publish it.

10-20M SE50bp reads is what I tell people I work with.

[bruce01]

Would LOVE to see the paper you refer to James!

[chadn737]

If you are working on a well sequenced genome and can map to a reference, then there really is no need to do paired-end if your only goal is differential expression.

[vehuardo]

Thanks for your quick response, James. Thats very informative. Surprisingly hard to come by any clear advice or common views on this topic.

Looking forward to the paper!

Cheers

[rcrystal]

Hi- the paper that introduces RSEM (Li and Dewey(2011). RSEM: accurate transcript quantification from RNA-Seq data with or without a reference genome. BMC Bioinformatics. 12: 323.) suggests that large amounts of single-end reads are the best for accurate gene level abundance estimates. However, if you're looking at isoforms, they recommend paired-end reads.

Apologies for not linking to the paper, first post, not sure how everything works yet!

[whataBamBam]

I have some data from PE Rna-Seq which was assembled de novo. When the guys at the sequencing centre aligned the reads back to the transcripts they just concatenated the fastq files together and aligned it as single end.

I can see this because I can see in the bam file the command they used to run the alignment. Does anyone know why they might have done this?

I have seen threads on here where people have suggested that if poor mapping is observed when mapping as paired end to try just concatenating the reads together and aligning as single end.