Preferred method for differential gene expression analysis?

[LeonDK]

Hi all,

So based on [1] I have implemented edgeR, except using the robust method of estimating the dispersion [2].

However having previously worked with non-parametric statistical methodology my attention has fallen on [3].

I am new to the whole NGS/RNASeq world and therefore it would be great to get some input reg. which method you prefer for differential gene expression analysis and why?

Cheers,
Leon

1. http://www.ncbi.nlm.nih.gov/pubmed/23975260
2. http://www.ncbi.nlm.nih.gov/pubmed/24753412
3. http://www.ncbi.nlm.nih.gov/pubmed/23981227

[bastianwur]

Review: http://www.ncbi.nlm.nih.gov/pubmed/24020486
Slightly related review: http://www.ncbi.nlm.nih.gov/pubmed/22988256

Personally: I use cuffdiff, mainly because it works without a hazzle, is fast, and doesn't require much additional effort for the input.
And I don't know R ^^.

[LeonDK]

Review: http://www.ncbi.nlm.nih.gov/pubmed/24020486
Slightly related review: http://www.ncbi.nlm.nih.gov/pubmed/22988256

Personally: I use cuffdiff, mainly because it works without a hazzle, is fast, and doesn't require much additional effort for the input.
And I don't know R ^^.

Hi bastianwur,

Thanks for the references, it seems that the consensus is that no method is better than another, each having its unique strengths and weaknesses.

Which basically means that scientists are likely to choose the one they "like" for whatever reason, ease of use, prior experience etc... Not a very scientific approach imho...

Cheers,
Leon

[LeonDK]

No offence btw. it's just frustrating not being able to get clear cut answer reg. which method is the better

Cheers,
Leon

[kopi-o]

That's because there *is* no clear cut answer

- If you have a complex design (more than one experimental factor varies), you want to use limma, edgeR or DESeq2. All reviews tend to agree that these are OK. I tend to favor DESeq2.

- Most reviews agree that CuffDiff is pretty bad.

- When you have many biological replicates, SAMSeq (a nonparametric method) is a good alternative.

- Ballgown can do DE analysis on novel transcripts and isoforms (because it's run on an assembly).

- Single-cell RNA-seq needs special considerations.

And so on ...

[LeonDK]

That's because there *is* no clear cut answer

- If you have a complex design (more than one experimental factor varies), you want to use limma, edgeR or DESeq2. All reviews tend to agree that these are OK. I tend to favor DESeq2.

- Most reviews agree that CuffDiff is pretty bad.

- When you have many biological replicates, SAMSeq (a nonparametric method) is a good alternative.

- Ballgown can do DE analysis on novel transcripts and isoforms (because it's run on an assembly).

- Single-cell RNA-seq needs special considerations.

And so on ...

I understand and accept that "there *is* no clear cut answer", my acceptance however does not dampen my frustrations

My setup is the following 96 RNA-seq sample run on Illumina HiSeq2000 using the Illumina TruSeq Stranded mRNA Sample Prep Kit. ~1 case per 4 controls.

I have done autmated QC'ing using Trim Galore! followed by mapping to UCSC hg19 using TopHat2 and then counted mapped reads using HTSeq. Each raw fastq-file contain ~60-80 mio. reads.

My current analysis of differentially expressed genes have been performed using edgeR Robust by Zhou et al. (10.1093/nar/gku310) and the workflow described by Anders et al. (doi:10.1038/nprot.2013.099)

Newbie here... Am I all good or do you have "crucial" input? ...and could you elaborate on the "complex" versus "simple" design?

Cheers,
Leon

[bastianwur]

No offence btw. it's just frustrating not being able to get clear cut answer reg. which method is the better

Cheers,
Leon

No offence taken, because as kopi-o said: There's no science yet in that part.
Just use the things which work for you and which give you the results you want (yes, I'm maybe naive, lazy, and a bad scientist, but well...in that case I'm fine with it).

Workflow seems so far normal, incorporates everything (no idea if trimgalore also does adapter trimming, but if so: good)...wait...besides rRNA filtering.
That seems to be missing.

complex design probably means time series and complicated relations between the conditions. No idea what your 96 samples are, but it probably qualifies.

[dpryan]

(no idea if trimgalore also does adapter trimming, but if so: good)

It trims adapters.

[kopi-o]

Newbie here... Am I all good or do you have "crucial" input? ...and could you elaborate on the "complex" versus "simple" design?

I think your approach seems sound.

By "complex" design I mean that there is more than one experimental factor that varies. For example, let's say you are looking at RNA-seq of tumor and paired normal tissue samples in several individuals. You *could* just compare the tumor vs normal groups, but you would get more statistical power by also considering which patient each sample is from - in other words, you'd model two factors, "individual" and "tumor/normal". This particular case would be a paired design ("paired" because you have "paired" tumor and normal samples from the same patient). This can be done in edgeR, limma, DESeq2, and in fact SAMSeq as well. If you do not use a paired analysis, natural variation between individuals can easily overwhelm the specific signal from the tumor vs tissue differences.

A more complex design could be a case where you have tumor cultures that have been treated with 3 different drugs at 3 different time points, with matched normals. Here you would perhaps want to model three different factors. It's this kind of scenario that you really need edgeR/DESeq2/limma for.

[LeonDK]

I think your approach seems sound.

By "complex" design I mean that there is more than one experimental factor that varies. For example, let's say you are looking at RNA-seq of tumor and paired normal tissue samples in several individuals. You *could* just compare the tumor vs normal groups, but you would get more statistical power by also considering which patient each sample is from - in other words, you'd model two factors, "individual" and "tumor/normal". This particular case would be a paired design ("paired" because you have "paired" tumor and normal samples from the same patient). This can be done in edgeR, limma, DESeq2, and in fact SAMSeq as well. If you do not use a paired analysis, natural variation between individuals can easily overwhelm the specific signal from the tumor vs tissue differences.

A more complex design could be a case where you have tumor cultures that have been treated with 3 different drugs at 3 different time points, with matched normals. Here you would perhaps want to model three different factors. It's this kind of scenario that you really need edgeR/DESeq2/limma for.

Super description - Appreciate it!

My setup is RNA-seq data from two groups: Group A from sick individuals and group B from healthy individuals and then I want to profile any gene expression differences

Cheers,
Leon

[dpryan]

Keep in mind that batch effects (they're inevitable) are easily handled by more complex designs as well. You also might benefit by stratifying patients & controls by gender or ethnicity or age or ... . An initially simple design can quickly become more complicated if someone forgot to take care of a factor during sample acquisition.