I have a question regarding analysis of a common experiment in regulatory evolution using DESeq2. The aim of the experiment is to distinguish cis and trans regulatory divergence between close species, and it is widely used on many organisms. The trick is that we profile gene expression in the two species and their F1 hybrid. We then determine allele-specific expression in the hybrid based on sequence differences between the two genomes. Differential expression between alleles of the hybrid results from mutations in cis, while genes that are differentially expressed between the parents and the difference is buffered in the hybrid, result from a mutation in trans.

I am working on two yeast species that in short are referred as C and P. I sequenced RNA of multiple samples of C, P and their F1 hybrid. I determined allele-specific expression in the hybrid and got counts for the transcript of the C genome (HC) and the P genome (HP) of the hybrid. Then I generated a “summerizedExperiment” table the differentiate genetic-background (“gb”) into four categories: C, P, HC, HP. I ran DESeq on the count table, using the design = ~gb. The contrast of C vs. P results in the log2 fold change b/w parents: A = log2(C/P). The contrast HC vs. HP is the log2 fold change b/w hybrid alleles: B = log2(HC/HP). The last is also referred as the cis effect, and genes the significantly change in this contrast could be referred are diverged in cis. The trans effect would be: A-B. My question is: can I use DESeq to do statistics on the trans effect (Null hypothesis: A-B=0, alternative hypothesis: A-B!=0)? What I understand is that DESeq performs a negative binomial test through fitting a GLM on counts data. The trans effect, as I defined, is the difference of two log2fold-change vectors. Can I somehow use the baseMean to perform the nbinomWaldTest on the trans effect (A-B matrix)?

Alternatives:

1. I tried an alternative of a t-test on the log2 fold change matrix, but it gives a very different result from DESeq and therefore not a good alternative.

2. I thought of an alternative: if my summarizedExperiment would look like: d = data.frame(sample=c("C", "P", "HC", "HP"), sp=c("C", "P", "C", "P"), F=c(0,0,1,1)) Where F is the generation (F0 is parent, F1 is hybrid). In this case I’m not sure what design and contrast should I use to get the cis and trans effects. Thanks in advance for your answer!

Just to generalize, if you want to compare two LFCs you can use interaction terms.

See if the information here helps:

https://support.bioconductor.org/p/61509/