I have been working on the analysis of time course microarray data (Agilent, Affymetrix, Illumina and Caltech platforms). I used Limma and Timecourse packages in R till now. I would like to know if there any other methods/tools/ideas on how to select differentially expressed genes from the set based on their time course expression profile. Some genes may be linear (increase or decrease) in their expression as the time increases and some genes can show a cyclic behaviour. Hence, What can be the strategies to choose differentially expressed genes?
We used the method of fitting Gaussian processes, and testing where in the timeseries a single Gaussian process best explains the data, or where two separate processes best explain. If two processes can explain the observed data better than a single one then we may call differential expression. I hope I have explained the principle (I did not personally do this work) but you can see the method used in our paper http://www.plantcell.org/content/24/9/3530.long and the reference to the original method from Stegle, 2010. There is also a Bayesian method BATS we have played with, which also works, but I don't think it has been recently maintained.
STEM is a freely available tool that can be used to analyze time series microarray experiments, where gene clusters with particular expression profiles can be selected and comparisons between treatments can be done:
http://www.biomedcentral.com/content/pdf/1471-2105-7-191.pdf
I have also found that self-organizing maps (SOM) are a very useful tool to analyze time series experiments, and makes it easier to visualize clusters based on their expression profile, which is very helpful when studying co-expression. Here is an example: http://www.jbc.org/content/280/27/25590.long (here transcriptomics and metabolomics data were combined, but can be used for transcriptomics alone)
The SOM tool is also free: http://prime.psc.riken.jp/?action=blsom_index
I thought there is a problem that limma treats each time point as independent whereas in a timeseries this assumption is violated because expression at time t+1 is dependent partly on expression at time t.
Hi all....thanks for your suggestions....will try all of them....i agree with Jonathan that Limma cant be used to infer/identify differential expression along the time series as we can only compare two time points at a time.....it considers the variation among the time points which we add i9nto our design matrix. In my view, Bayesian and Gaussian approaches would suit more favourably.
We previously analysed a gene expression time series using limma and compared every time point to time point zero to derive foldchange values. It depends on the question you have, I suppose. You could also try out a cluster analysis based on the gene expression levels themselves. Maybe have a look at MEV (http://www.tm4.org/mev/). Its a suite with lots of builtin functionality.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Bioinformatics
- Bioinformatic Tools