SOM (Self-organising Maps) to identify expression trends in time-course data

[rpauly]

Hi,

I am curious to know if anyone has employed SOM (Self-organising Maps) to identify expression trends in time-course data (RNASeq/Exome seq)?

~Thanks,
Rini

[dpryan]

There are a few papers that use SOM with RNAseq in general (a time-course exome-seq experiment would rarely make any sense), though I don't recall that they use it in the context of a time-course experiment (there's no reason that wouldn't work though). Just search pubmed for them if all you need are some papers.

[rpauly]

Thank you for the quick reply!
But I do not see so many papers with SOM and RNASEQ (maybe 3?).

Do you have specific references that you have come across?

~Thanks!

[dpryan]

There aren't that many since it's not a terribly popular method. Also, there are probably some papers using it on microarrays (the same concepts will apply).

[Dario1984]

There are better statistical modelling approaches available, such as GPclust. It allows biological replicates to be properly used. It's made for normally distributed data, but you could transform your counts to be normally distributed.

[sarvidsson]

There are better statistical modelling approaches available, such as GPclust. It allows biological replicates to be properly used. It's made for normally distributed data, but you could transform your counts to be normally distributed.

Not wanting to hijack the thread, but I'd be interested in trying GPclust on one large RNASeq dataset I'm working on currently. Did you use it, and if so, is it worth spending time to try it out?

[JamesHensman]

GPclust

GPclust works well for me, but then I'm the author

You can find the code and some demo IPython notebooks here http://staffwww.dcs.sheffield.ac.uk/...n/gpclust.html

If your data is cleanish, GPclust can provide nice results like this one

https://drive.google.com/file/d/0Bz7vcLVtU_zoZ05FbGVoMnlqUVU/view?usp=sharing

Apologies for the self promotion -- I'm happy to help if there are other questions.

[sarvidsson]

GPclust works well for me, but then I'm the author

You can find the code and some demo IPython notebooks here http://staffwww.dcs.sheffield.ac.uk/...n/gpclust.html

If your data is cleanish, GPclust can provide nice results like this one

https://drive.google.com/file/d/0Bz7vcLVtU_zoZ05FbGVoMnlqUVU/view?usp=sharing

Apologies for the self promotion -- I'm happy to help if there are other questions.

I'll try it out then. How would you define "cleanish"? My data is a large (200+ samples) and rather deep (30 Million+ reads per sample) expression set aligned to a draft transcriptome assembly, which is not completely cleaned up (i.e. there will be some redudancy in the transcontigs and some contaminating species transcripts). Also, some of the sample points/replicates had problems (degraded RNA, contamination from other species etc.) so they will introduce some errors (the worst samples were selected out, however).

No offense, but my experience is that "clean" data is the exception, and is mostly encountered as example datasets in bioinformatic publications describing analysis methods I'd be interested in a method which is robust to the problems described above.

[JamesHensman]

I'd say your data needs to be free of outliers and other nasty behaviour.

I've taken to filtering signals for signal-to-noise ratio, by dividing the variance of the replicate means by the mean of the replicate variances. You can still cluster 1000s of genes with gpclust, by too many genes which are just noise will confuse it.

I would say that a method that deals with the problems you describe probably depends mostly on good data munging, rather than the method itself.

[sarvidsson]

I'd say your data needs to be free of outliers and other nasty behaviour.

By (partly manually) looking (PCA, time plots...) at some 20 genes with known behaviour, we've removed obvious outlier samples.

I've taken to filtering signals for signal-to-noise ratio, by dividing the variance of the replicate means by the mean of the replicate variances. You can still cluster 1000s of genes with gpclust, by too many genes which are just noise will confuse it.

By "too many genes which are just noise" do you mean that they show erratic behaviour over time or over replicate - or both? I've looked at a lot of expression data over the last few years, and typically most genes show erratic behaviour in either respect - be it due to uncontrolled biological or environmental variation, imperfect replication or whatever technical difficulties thereafter.

I would say that a method that deals with the problems you describe probably depends mostly on good data munging, rather than the method itself.

Agreed - but be careful in distancing your method too far from the application of it. Most "successful" (widely used) bioinformatic packages (e.g. for variant detection or expression analysis) include documentation and hands-on examples (using published, real datasets with all kinds of biases and noise, not only simulated data) on recommended practices for raw data pre-processing, normalization, filtering of "noisy" samples, filtering of "noisy" genes etc... While I understand that researchers do not always have the time for maintaining such documentation or providing support to users, IMO this is a key to "success" (in the sense of getting a well cited paper).
I must admit I didn't read your IPython notebooks yet, so I better do that now...

[sarvidsson]

So I've checked your IPython notebooks, and the examples make sense to me. The Kalinka dataset is microarray data on log2-scale - correct? So I should be able to use RNA-Seq count data processed with DESeq2's rlog or vst, right? (http://www.bioconductor.org/packages....pdf#section.2) I guess I should skip your normalization step then, however...

You might find that more novices (at least here on SEQanswers) would want a walkthrough for data from an RNA-Seq experiment, e.g. from an unprocessed count table from HTSeq-count.

I could try this the hard way, but would it be feasible to cluster >=10 000 genes in reasonable time (given that they are well filtered)? If not, are there steps in the algorithm that could be parallelized to achieve that?

[Dario1984]

Did you use it, and if so, is it worth spending time to try it out?

I haven't tried it.