We just finished the QC for the total reads obtain from a Time Course RNA-Seq experiment (Illumina HiSeq2000), and are satisfied with the quality.
Now we have started the Differential Expression Gene analysis, and the strategy proposed by our bioinformatican was the following:
Calculating DEG for every condition compared with the initial one, so we will have at the end 5 comparisons. DEG for each gene will show a certain p-value in each case, so we would count as statistically significant those genes with p-value<0.05 in all 5 comparisons.
What is the main goal here ? Do you also have the control for each time point or do you just want to see the genes that are altered in each time point w.r.t 0.h ?
If you are looking for genes that follow a pattern, you could use something like WGCNA to get the tightly co-expressed genes across all time points but you need to have enough samples.
You can also use JTK_CYCLE to get the cycling genes. I heard about this program from this science paper.
If you do pair-wise comparisons is that going to be done with just one sample each or one sample to some replicates of control or are there replicates for all samples at all time points?
Sounds about right....but there will be more considerations that the bioinformaticioan should know about...replicates etc.
Take a look at the original paper for cufflinks. They test they're new software on cells going through different stages of differentiation so gather data at 4 time points (instead of your 5). How they analyse and present the data should give you more ideas. The bioinformatician will (should!) already be familiar with cufflinks.
Agreed to all advices above... If you have enough replicates for each time point, you may also want to do clustering for differentially expressed genes in each time point.
My understanding is that your recommended TimeCourse package seems to have been designed for microarray data, and so loosely - the voomtransform in limma is to ensure RNA-Seq data can be modified to comply with the statistical assumptions underlying microarray data and its analyses, right?
Given that it's been almost 4 years since the original post here, my question is whether limma + timecourse would still be the best or one of the best ways to analyze time series data today. Your thoughts / advice? TIA!
This is certainly an old post, however the approach I outlined is still valid. I'd spend some time thinking about what your experimental question is based on time. Another approach might be fit a cubic spline curve - see section 9.6.2 of the limma users guide.
Mostly the best statistical approach and testing is reliant on what your experimental question is.
Approach looks correct. It appears that you have six time-points, and therefore you will have 5 comparisons with respect to the first time point.
I would recommend you to use adjusted p-value (also called q-value) < 0.05. Additionally, you can apply some filter on fold-change value (like 2) also.
I am not sure about your pipeline. But if you have count data, you can use DESeq2 on your time-series experiment. There are different approaches to analyze time-series data based on your experiment [Reference: http://bioconductor.org/packages/release/bioc/vignettes/DESeq2/inst/doc/DESeq2.pdf].
You can also look at Enriched pathways using GSEA (Gene set Enrichment Analysis) after doing Differential Expression using DEseq package.
There are lot of Categories in GSEA (Oncogenic,Immunologic etc) or you can curate your own signatures and do your enrichment analysis.
Another thing you could do is do heatmaps between the categories using differentially expressed genes(from visualization perspective)
What is the main goal here ? Do you also have the control for each time point or do you just want to see the genes that are altered in each time point w.r.t 0.h ?
If you are looking for genes that follow a pattern, you could use something like WGCNA to get the tightly co-expressed genes across all time points but you need to have enough samples.
You can also use JTK_CYCLE to get the cycling genes. I heard about this program from this science paper.
If you do pair-wise comparisons is that going to be done with just one sample each or one sample to some replicates of control or are there replicates for all samples at all time points?
Sounds about right....but there will be more considerations that the bioinformaticioan should know about...replicates etc.
Take a look at the original paper for cufflinks. They test they're new software on cells going through different stages of differentiation so gather data at 4 time points (instead of your 5). How they analyse and present the data should give you more ideas. The bioinformatician will (should!) already be familiar with cufflinks.
Agreed to all advices above... If you have enough replicates for each time point, you may also want to do clustering for differentially expressed genes in each time point.