Is Fold Change reliable? Does anyone use Characteristic Direction?

[hubertofliege]

I'm taking a course on Coursera taught by Avi Ma'ayan. In one module, a member of his lab promotes an alternative method of determining differential expression called "Characteristic Direction". He then plots several comparisons of success rate of different methods. On his charts, fold change looks terrible.

Is this reliable? I don't see anyone else using this method he suggests, "Characteristic Direction". Can anyone set me straight?

https://bmcbioinformatics.biomedcentral.com/articles/10.1186/1471-2105-15-79

Thanks

[dpryan]

The fact that it relies on covariance matrices reminds me of WGCNA. In practice, the normal methods are reliable enough and are somewhat easier to understand and debug than something that's based on LDA. There's also the reality that the L1 regularization is probably going to get rid of a LOT of genes that inevitably someone is going to ask about those. There's also the fact that if one uses the normal methods then the whole problem is no longer ill posed...

[dariober]

I don't see anyone else using this method

In addition to Devon's insightful answer, I would add a couple of possible reasons:

* limma, edger and DESeq have been used very extensively and they have very good reputation (well deserved). Their documentation is great and questions about their usage gets answered quickly on mailing lists. For these reasons It's difficult for a new method to occupy their niche "just" by promising slightly better sensitivity and at the cost of what Devon points out.

* It seems to me that the characteristic direction method at the moment can cope only with two conditions, one vs another like a t-test. limma, edger and DESeq can easily accommodate more complex designs.