Venn diagram type analysis of ~6 conditions

[Noa]

I have RNASeq data of five different treatments (some with two time points and some with one). I am trying to think of a good visualization. Is there any tool out there to do a Venn diagram type figure of general number of up/downregulated genes to see which are shared by >1 treatment?

I am also happy for any input on how to attack such a massive amount of data (where to put the cutoff - I have been using 1.5 fold expression difference from control, FPKM of at least 10) and how to do pathway enrichment analyses, etc.

Thanks!

[jmw86069]

There are other threads on Venn diagrams, but I think the short of it is that once you get beyond 4 sets (maybe even just 3), unless the visual message is blatantly obvious it's difficult to convey with that type of diagram. However, if you did want to try >4 circles, I'm impressed with the R package venneuler by Dr. Lee Wilkinson and Dr. Simon Urbanek. Some snapshots are shown here:

http://rgm2.lab.nig.ac.jp/RGM2/func.php?rd_id=venneuler:venneuler

What impresses me about this package is not just that it handles proportional size and overlap of circles, but it provides a stress factor which in essence scores how well it fits the data. Dr. Wilkinson's publication compares to other methods, and it strikes me his is one of the only tools that scores the goodness of fit at all. When you have more than 4 circles, there's no guarantee you can ever visually present it perfectly, however their stress factor at least tells you how close it is.

That said, a heatmap of fold changes would probably be visually informative. Do the classic hierarchical clustering of columns/rows with Cluster-TreeView, or any number of R functions like heatmap.2(), and it may be clear. The standard trick of filtering for significant results before clustering will probably help clean up the noise. But with that many comparisons, one suggestion is not to think of it too much like one giant dataset, to the detriment of focusing carefully enough on each comparison.

For cutoffs, hard to suggest much without evaluating the variance, applying some statistical model, etc. Using DESeq and/or edgeR or their extensions (e.g. DEXSeq for exon-specific changes) will help establish the level of confidence you may have with the data. But as DESeq author Dr. Anders suggests, a heatmap of fold changes may not work as well for counts data as it would for 'omics platform data. Depends upon your data though.