Forum:"Separate enrichment analysis of pathways for up- and downregulated genes"

Question

[Woa]

I recently came across this paper. I wish to know if anybody already following this method or not.

Separate enrichment analysis of pathways for up- and downregulated genes

Guini Hong, Wenjing Zhang, Hongdong Li, Xiaopei Shen and Zheng Guo

J. R. Soc. Interface 6 March 2014 vol. 11 no. 92

http://rsif.royalsocietypublishing.org/content/11/92/20130950.short?rss=1

Abstract:

Two strategies are often adopted for enrichment analysis of pathways: the analysis of all differentially expressed (DE) genes together or the analysis of up- and downregulated genes separately. However, few studies have examined the rationales of these enrichment analysis strategies. Using both microarray and RNA-seq data, we show that gene pairs with functional links in pathways tended to have positively correlated expression levels, which could result in an imbalance between the up- and downregulated genes in particular pathways. We then show that the imbalance could greatly reduce the statistical power for finding disease-associated pathways through the analysis of all-DE genes. Further, using gene expression profiles from five types of tumours, we illustrate that the separate analysis of up- and downregulated genes could identify more pathways that are really pertinent to phenotypic difference. In conclusion, analysing up- and downregulated genes separately is more powerful than analysing all of the DE genes together.

[mforde84]

We filter like this a lot for certain down stream analyses. For example, GO term enrichment analysis is a good example. Seems to me that were just talking about the differences between one and two tailed significance testing. Two tailed is going to be more conservative because the threshold pvalue is split to both sides of the NULL distribution, whereas one sided cutoff is for just one side of the distrubion. While this might make it easier to call significance in one direction, that's really all it does. You could have the same effect by doing a two tailed test, and setting you're pvalue cutoff to 0.1. Unless you can justify why you're doing a one tailed test, then you shouldn't be doing a one tailed test.

Answer 1

[mikhail.shugay]

Interesting, but it's behind the paywall. Hope they've shown that while increasing statistical power thay don't increase false-positive rate. Otherwise seems to be the same as to state "two-tailed P-values tend to be higher than one-tailed" :))

[mikhail.shugay]

PS. There is of course a vast field for debates. For example consider JAK2/STAT5 pathway, known to be involved in ocogenesis. There are both "positive" regulators of pathway activity, like EPOR/JAK2/STAT5 and "negative" ones like SOCS/PIAS genes. So if up-regulated gene group contains "positive" regulators and down-regulated one contains "negative" regulators the pathway is generally regulated. Of course on general basis you usually cluster genes with similar expression profiles and annotate these clusters, but it could be sometimes more wise to take into account positive/negative interactions between genes..

[Istvan Albert]

or another interpretation could be that by splitting the data they end up making fewer comparisons, thus the correction for multiple testing is smaller.

Answer 2

[Irsan]

Like many others, I always make a heatmap of the expression estimates of all differentially expressed genes arranged by hierarchical clustering. Then I perform ontology analyses on correlated sets of genes.

Answer 3

[Charles Warden]

I always analyze up- and down-regulated gene lists separately: I think that makes the interpretation easier.

That said, there can still be mixed results. For example, the KEGG Wnt pathway includes both agonists and antagonists, and I did have one circumstance where this was an issue. This is why I developed BD-Func.

https://peerj.com/articles/159/

However, I admittedly think standard analysis of the up- and down-regulated genes is typically OK.

Answer 4

[ankita0007]

I am trying to differentiate treated and control conditions on the basis their biological process, molecular functions and cellular components. So far, I have performed abundance analysis with the help of htseq. Thereafter, differential expression analysis with the help of limma package. Presently, I am having the chunks of differentially expressed gene sets having p < 0.05 and fdr < .05 falling in different categories and now I want to perform separate pathway analysis of up and down regulated genes falling in different categories separately which I have got it by fitting the contrast and performing decide test on it.

Before thinking about doing the pathway analysis separately, I have tried to do the same with camera but there the problem was: 1. It does not separate up and down regulated genes of different categories and perform the analysis of each contrast. 2. Also, by performing the analysis of each contrast I am not getting the appropriate fdr.

So, I think for now it would be the great help if you would suggest me how can I go ahead with the pathway analysis of different chunks of the gene sets. Also, excuse me if somehow I am not making sense as I am a beginner in the field.

Answer 5

[theobroma22]

You can simply use SPIA with all genes and the output will tell you if the pathway is activated or inhibited, which is basically the same as separating them. Also, I would be recommend caution when choosing your sources of information or who you choose to cite.