Neurons were treated with four different types of drugs, and then a full transcriptome was produced. I am interested in looking at the effects of these drugs on two specific pathways, each with around 20 genes. Would it be appropriate for me to just set up a simple comparative test (like a t-test) and run it for each gene? Or should I still use a differential gene expression package like DESeq2, even though only a few genes are going to be analysed? The aim of my experiment is a very targeted analysis, with the hopes that I may be able to uncover interesting relationships by cutting out the noise (i.e., the rest of the genes that are not of interest).
Hello! It would help to know what you mean by full transcriptome was produced? Are you doing bulk RNA-seq with a relative comparison experiment for each drug with biological replicates? Do you have the dreaded singleton RNA-seq in a profiling experiment? These are bad, not what anyone should be encouraged to analyze. What I would do is read the "what to do if you have no replicates" section in the edgeR users guide. Additionally, I would include data from the whole transcriptome measurement using "model based normalization" with contrast matrices to pull out the differences. I think if you were to just make a heatmap from the subset of genes, you would want to make RPKM to account for the differences in total reads between the libraries. After you use edgeR to get the log-fold change and FDR p-value you could then use directed analysis with QIAgen IPA if your university has institutional access. Good luck!!
https://bioconductor.org/packages/release/bioc/vignettes/edgeR/inst/doc/edgeRUsersGuide.pdf
The full information should be used for normalization and for a robust variance estimate. After this you might focus on a (small) subset of genes that are of interest to you. The subset must be defined prior to the experiment (it's selection must be independent of the results obtained from that experiment; otherwise you use information twice, leading to "self-fulfilling prophecies" rather than to rigrous tests of meaningful hypotheses). If you screen for candidated within the selected pathways, the p-values would have to be adjusted for multiple testing for the actual number of tests performed (30-40, as you said). However, I don't think that this is a smart approach. Your research hypothesis indicates that the orchestrated effects of the genes are relevant, not the effects of single players. Hence, looking for individual statistical significance of single genes may not be adressing your research hypothesis. Instead I think you should model the (modulation of) signal transduction and response according to the observed changes in gene expression (of all genes in the respective pathway) and see if this supports or contradicts your research hypothesis. This is not a quick-and-easy thing to do and required a lot of background knowledge, understanding and commitment, but it adresses (as I think) the relevant questions. The easy way is to throw some t-tests or whatever on the data adressing irrelevant questions...
Matthew Edward Thornton Thank you for your answer! To clarify, there are six groups (four drug groups and two controls), each with four replicates. I have been supplied with a raw gene matrix with the expression of ~58000 genes. I am only interested in circadian genes (I think there's actually 30-40 genes of interest in that particular group, my mistake) and genes relating to HPA axis activity. My plan was to normalise the whole dataset first (I was recommended TMM, but can do RPKM instead) and then do my analysis. I thought targeted analysis of only these pathways would be ideal as it might cut out the noise; you're saying I could use EdgeR for this purpose? I've checked and I do have access to IPA so I'll definitely use it if you think it's a good idea!
Jochen Wilhelm Thank you for your answer! I've supplied a bit more information about the data in a previous response. I think that looking at all genes in the respective pathway would be a good idea. How would you recommend going about this?
Heather Macpherson that is much better! I think edgeR or limma would be highly appropriate to process your data. I would just use the raw count data. You should should only import raw count data into edgeR. Limma is an exception because it was developed for microarrays which are normalized intensities. Still its not good to do this. The edgeR and limma user guide is an excellent resource and has many tutorials on its proper use. as Jochen Wilhelm explained very well you will not want to subset. In edgeR and limma you can filter by experiment which would require a design matrix. i would also generate a contrast matrix for the group comparisons. After your groupwise comparisons you can subset as you like. highlight those genes in a volcano plot or smear plot. If you have not done this before, I would highly suggest starting from homer step 1 and then directly to the edgeR user guide vignette. http://homer.ucsd.edu/homer/basicTutorial/index.html
So with IPA after a "core analysis" of your groupwise comparisons from edgeR, you can make a gene list with the genes of interest and then add them to a "new network" and "connect" the network which will pull in more genes and then you can prune and grow branches until you have a complete network. You can also color the network with your experimental data by log-fold change & FDR P.Value from the different groupwise comparisons. These networks you can also export from IPA into two files which you can merge with your raw data and then modify in Cytoscape. You have to be careful with the IPA knowledge base as it is not always current and correct, but it is an excellent and quick and easy starting point. Good luck!!
- Badges
- Science topic
- Similar topics
- Biological Science
- Microbiology