Time-Course Experiment - What Statistical Test To Use?

Question

[vanzetti]

Hello.

This is my problem. I have data from label-free proteomics experiment. I have 4 strains with 5 time points in each and 3 replicates. One of the strains is a negative control and expresses nothing, while 3 others are induced to express different proteins at time point 0. Overall I have complete data (all strains, time points and replicates) for around 1000 proteins.

What I'm interested in is finding which genes have significantly different profiles of expression in the other 3 strains compared to negative control. And I'm not sure what to do? Should I just make a t-test of negative control vs. each of the strains in all 5 time points? Or is there a more sophisticated solution?

Thanks.

[Tky]

There are a number of software/packages for time series microarray experiments, many based on clustering methods like hierarchical clustering. Please do a Google search for the detailed methods. You may try your data to software like STEM: a tool for the analysis of short time series gene expression data. https://www.cs.cmu.edu/~jernst/stem/

[vanzetti]

The problem in my case is how to discover what is significant, not how to cluster. But thanks anyway.

Answer 1

[Charles Warden]

You can try EDGE for time-series analysis (and I think SAM also has a time-series module):

http://www.genomine.org/edge/index.html

I have also used the 'timecourse' package in R and found it to be quite useful:

http://www.bioconductor.org/packages/2.12/bioc/html/timecourse.html

[vanzetti]

Thanks, EDGE seems partially useful for the task. The problem is, I'm comparing 4 strains, and it can't tell me which one is different compared to which one...

[Charles Warden]

True, but you could either compare each of the 3 strains to the control separately and/or pool the 3 strains compared to the control. I'm guessing that it isn't really important how much the 3 altered strains vary from each other (although you could continue to do pair-wise comparisons, if needed).

Answer 2

[Michael]

I think ANOVA is a better choice in this case, than an all versus all t-test which will inflate the type-I error due to multiple testing. If you suspect the data to deviate strongly from the assumptions, you can use the Kruskal-Wallis non-parametric alternative instead. Both methods should be available in R.

[vanzetti]

But with ANOVA I won't know which of the strains is significantly different from negative control. It can also miss the situation where Mut1 and Mut2 are significantly different from each other, but neither of them are significantly different from the negative control.

[Michael]

If you want to find out which group is contributing, you can also do a post hoc-test individually for genes coming up in the ANOVA. There is little bit of speculation involved in your second assumption, it depend on how you specify the model formula. But in general ANOVA should have more power than an all versus all t-test in such a setting.

Answer 3

[Steve Lianoglou]

The limma user's guide has a section on how to analyze timecourse data with linear models using splines, which might be of interest for you.

If you protein expression data comes from peptide counts, perhaps you'd want to shoot the data through voom (also explained in the limma user's guide) before doing the linear model fitting.

Answer 4

[Houkto]

Hi there,

I sometimes go for another approach which is a cluster analysis using BiolayoutExpress. Normalize your data and check the data format at their website which biolayout will accept then run MCL button and then view the clusters as plotted graph. Look for a a cluster with gene expression pattern match your hypothesis across your samples and time.