What Is A Good Threshold For Log2 Fold Change?

Question

[Some]

Hi there,

for a pilot study I have to find out differentially expressed genes between time-points of a microarray time-course study. There are only 4 chips and I am totally aware that this has not a lot of statistical power, it is basically fishing.

My approach is to calculate the log2 fold change between array X and array Y. What would be a good threshold for that please?

Thank you very much.

Answer 1

[Lars Juhl Jensen]

Any fold-change threshold would be completely arbitrary. What I would normally do in a situation like this is to first rank all genes based on their (absolute log) fold-change. I would then look at the top ranking genes and verify that these indeed make sense.

Next, I would try to get hold of some sort of gene set that I can use to more systematically assess if the ranking makes sense. For example, if the array study has to do with a certain process, I would try to get hold of set of genes that are known/believed to be involved in the process. Depending on the experiment, this list could come from a pathway database like KEGG, a disease database like OMIM, or from text mining of PubMed. Once I have such a reference set, I can plot sensitivity (i.e. the fraction of genes from the set that are found) as function of the rank of the gene.

Why would I make such a plot? Firstly, it allows me to assess if one way of ranking the genes is better than another; maybe ranking the expression profiles based on something else than the maximal fold-change is better. Secondly, and directly related to your question, it allows me to select a cutoff that is not completely arbitrary: by looking at the curve, you can tell how long a list of "differentially expressed" genes it makes sense to propose based on the experiment.

Answer 2

[User 59]

So are there 4 time points on 4 chips? If so then there is no statistical power.

I think most people, out of a sense of tradition or habit (if nothing else) would normally look at a 2-fold change for an unreplicated experiment. However you will need to spend time following up any leads via another quantitative method. You recognise this is a fishing exercise, so as long as you remember to treat it as such...

The point is any fold change cut off is entirely arbitrary, so will largely be set by the number of targets you wish to follow up as as result of this work, and whatever knowledge of the biological system you have that you can bring to bear on the analysis.

Answer 3

[Panagiotis Alexiou]

In this type of questions and when rigor is not expected, I usually use thresholds over 1 and under -1 of the log2fc. This means genes that doubled or halved.

Answer 4

[Stefano Berri]

I second Daniel Swan

also, fold change depends on intensity (if you don't have a p-value), as there is more vairability at low intensities, so always keep an eye on the intensity as well as the fold change. Sort them by fold change, but have the average intensity on the side. A fold change of 3 but very low intensity is not very reliable...

Answer 5

[User 5154]

Hi all,

Is it possible to do a t test for checking the significance of a particular gene's expression by comparing the log2 fold ratios from untreated ( 5 time points) and treated( 5 time points)? Unreplicated samples and I am wondering whether a t test is valid with unreplicated time course?

Thanks! Sabna

[Michael]

A test is not possible because there are no replicates, unfortunately.

Answer 6

[Sabna]

My samples are not that far apart.Its just 30 mins, 1 hours, 2 hours, 4 hours and 6 hours post treatment. Across this total 6 hours genes dont change much as shown by the log2 ratios across each time point. The question is, could I take each of those time points as biological replicates and do a t test between conrtol and treated?

[Michael]

Are you the original poster (Some)? Than you have used multiple accounts for posting (Some, unknown, and Sabna), please stick to one account.