Is There Any Other Method For Isoform-Level Differential Expression Analysis Except Cufflinks/Cuffdiff?

Question

[Huanwei Wang]

Hello, all!

I'm a beginner for RNA-seq data analysis.

As far as I know, there are many tools for gene expression(such as cufflinks/HTSeq-Count/even bedtools for RPKM) and differential expression analysis(CuffDiff/DESeq/EdgeR) of RNA-seq. However, for isoform-level analysis, is there any other method except cufflinks/cuffdiff? And which method or pipeline do you think is the best?

Thank you!

Answer 1

[Charles Warden]

I have personally not been very impressed with the transcript-level qualifications, especially with RNA-Seq data with very uneven coverage (which is almost all the data I have directly worked with, although I've heard there are newer protocols to assist with the problem of uneven coverage). For this reason, I would tend to stick towards analyzing differential splicing events (exon skipping, intron retention, etc.) with MATS, MISO, etc. over whole transcript quantification (and I tend to use gene-level mRNA quantification rather than transcript-level mRNA quantification)

That said, you could take the transcript abundances from cufflinks, RSEM, etc. and treat them like a normal, gene-level differential expression experiment (using limma, sRAP, etc.). I don't think this a great solution, but I don't know if it is really much worse than using cuffdiff.

To get an idea about the robustness of gene-level vs. transcript-level differential expression would look like, you can see Figure 5 in the following paper (although it may not be a completely fair comparison because the gene-level and transcript-level expression will often be correlated, and discrepancies between transcript abundances shouldn't be expected for all genes):

http://bioinfo.aizeonpublishers.net/content/2013/6/bioinfo285-292.pdf

[Varun Gupta]

HI Charles,

Could you explain me the difference between gene-level mRNA quantification and transcript-level mRNA quantification

[Charles Warden]

The precise difference may vary depending what tool you use, but gene-level quantification will give you a single expression value per gene (associated with a gene symbol, most likely). In contrast, transcript-level quantification will give you expression values for each isoform of a gene, and this can only be calculated when using a tool that assigns reads across isoforms for a given gene (often associated with a RefSeq accession number, but that would obviously depend on what reference you used to define your genes).

I would define the gene-level quantification as the sum of reads assigned to exons present in any of the isoforms for a gene, but I'm not sure if this is always the case.

Answer 2

[Ryan Dale]

Have a look at MISO (paper). The documentation is quite good, too.

[Puriney]

That document is quite impressive

Answer 3

[Mikael Huss]

Yes, there is EBSeq, which is often coupled to RSEM as the isoform quantifier. The EBSeq paper was just published.

EDIT: I forgot about ALEXA-seq, which is a good choice for human and mouse RNA-seq.

Then there is the possibility to obtain isoform-level read counts, which could be done by CuffDiff, RSEM or eXpress (although the latter works by mapping to a reference transcriptome, so it is a bit different from the others), and to then use DESeq/edgeR/limma/SAMSeq ... (insert favorite DE tool). However, I have heard it argued that this is not correct because the DE tools build on assumptions about read counts that apply to genes but not isoforms. For example, the EBSeq paper abstract puts it in this way:

When isoform DE is of interest, investigators often apply gene-level (count-based) methods directly to estimates of isoform counts. Doing so is not recommended. In short, estimating isoform expression is relatively straightforward for some groups of isoforms, but more challenging for others. This results in estimation uncertainty that varies across isoform groups. Count-based methods were not designed to accommodate this varying uncertainty and consequently application of them for isoform inference results in reduced power for some classes of isoforms and increased false discoveries for others.

[DG]

I believe the newer DEXSeq (from the DESeq people) will no do isoforms because it evaluates differential expression at the level of exons and not genes.

[Huanwei Wang]

Can you compare the tools you mentoined with cufflink/cuffdiff?Which is more easy to use, more exact?

[Mikael Huss]

It's hard to compare. I find Cufflinks/Cuffdiff straightforward to use but that could be because I am more used to them. I have run into some practical issues running RSEM but ultimately got it, and EBSeq, to work fine. However, it is very hard to say which is more accurate because of course I do not have any ground truth for the experiment I was looking at. There is a good comparison of differential expression methods for RNA-seq (http://arxiv.org/abs/1301.5277) but it only looks at the gene level (which of course is a good first step), not the isoform level. It would be cool with an isoform-level version of that paper.

[mjg]

Hi Mikael,

Have you used EBSeq? I have paired/matched samples, and I need a tool that takes this factor into consideration. I have used the latter alternative you mention: estimating isoform expression with CuffDiff and then doing a paired test in edgeR.

However I would prefer to use a tool that does both expression estimate and DE. Do you know if EBSeq supports paired samples?

Thanks,

Maria

[Mikael Huss]

I don't think it supports paired samples (please correct me if I am wrong, someone). Right now I am waiting for Sleuth (https://groups.google.com/forum/#!topic/kallisto-sleuth-users/WxWLcmE1Eeo) which goes together with the Kallisto transcript-estimation software and which will include functionality for specifying designs (such as paired samples). That would allow for isoform-level expression estimation together with factorial design aware DE analysis. You could also try Ballgown, which starts from Cufflinks assemblies, estimates expression levels of contigs (i e transcripts) which you can then do paired DE analysis on. However, if you have a well annotated genome you will probably lose some sensitivity in this way.

Answer 4

[Ann]

Also give Sailfish a try.

See: http://www.cs.cmu.edu/~ckingsf/software/sailfish/

Link to recently published Sailfish paper: http://www.nature.com/nbt/journal/vaop/ncurrent/full/nbt.2862.html

[Rob]

Sailfish has been massively updated in the past month; the underlying algorithm has been improved considerably. The newest release is markedly faster, uses fewer resources, and is more accurate than the previous stable version. Some interesting new features (e.g. variance estimation of transcript expression via Gibbs sampling and bootstraping) have also been added. If you haven' tried out the new version, I recommend you give it a spin!

Answer 5

[fo3c]

Have you seen MMSEQ?

Haplotype and isoform specific expression estimation using multi-mapping RNA-seq reads

[Mikael Huss]

It is an interesting paper which I hadn't seen, but it seems to be about isoform level quantification and not about isoform level differential expression analysis as the question specifies

Answer 6

[vipin.ts]

I think rquant will perform isoform level quantification bioweb.me/rquant

Answer 7

[Leszek]

In ENCODE RNA-Seq pipeline they used Flux Capacitor to estimate isoform expression.
I had difficult time running it though.

[David Quigley]

Anyone evaluating Flux Capacitor (which I have not myself used) might be interested in this blog post from Lior Patcher on the subject.