Microarray Time Series With No Replicates

Question

[datanerd]

Hi all, I have a 17 point time series data with no replicates. I wanted to find significant gene profiles over this 17 time points. Is masigpro R package good for my data? This is just one sample treated after injury (i.e there is no treated vs control) Apart from Matlab, what other R packages would you suggest for such data? Thanks in advance.

Answer 1

[Sean Davis]

With gene expression data, you have essentially two large classes of analysis and you will likely want to apply both:

1. supervised
2. unsupervised

For supervised analysis, you might be able to make "pseudo-replicates" by treating your 17 time points as early, middle, and late, for example. Then, use any number of statistical packages to compute differentially-expressed genes between the groups. For unsupervised analyses, you have a lot of options including hierarchical clustering, k-means clustering, PCA, NMF, and many more. The trick to these is to limit to a set of genes that is regulated in your system. Often, using the top X% of the most variable genes (where X is chosen to keep the number of genes manageable) will suffice. Cluster only the genes and order the samples by time to see trends in the data.

[datanerd]

Thanks for the answer Sean. I was wondering with the "pseudo replicates", since my data is over four days, can I treat each day time points as replicates? And is this the same as what is said about the dummy variable setting in masigpro package for long time series?

[Sean Davis]

I cannot tell you what make biological sense in your system, but you could try each day as a separate factor level, yes. I do not know much about the masigpro package, so I cannot really comment on it directly.

[datanerd]

the data is the treated cells with a drug after every24 hours. But i found many genes to respond differently on the first day than the others. So even I am not sure if that makes biological sense to treat them as replicates. Would appreciate your insight on this. Thanks for the suggestions though :)

[Sean Davis]

Unfortunately, your experimental design is less than optimal. For supervised analyses, replication is REQUIRED. However, you can still do unsupervised analysis.

[datanerd]

Thanks. Will try unsupervised analysis.

Answer 2

[Charles Warden]

While I agree that it is always best to have replicates, I think timecourse will provide an option to analyze your data:

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

Answer 3

[JC]

I suggest you Limma. Because you have no controls per each time point, maybe you can use T0 as the control for all points.

[Sean Davis]

Limma will still not be very useful without replicates.

[JC]

I agree, it's always hard to do good analysis without replicates.