Simple stats question

[jvanleuven]

Hi all,

I have a pretty simple question on data anaysis, but I'm getting a bit lost in finding the correct statistical tests to run. The data looks like this:

Organism: Gene: Type: ExpressionFoldChange:

Bacteria1 Gene1 Metabolism 10
Bacteria1 Gene2 Metabolism 2.2
Bacteria2 Gene1 Metabolism 1.0
Bacteria2 Gene3 VitaminSynth 0.2
....
....
....

I want to identify variables that are associated with the greatest fold change.
I have two questions: 1)what is the best statistical test and R function to do this, and 2)do i need to normalize the fold change data somehow? the range is from 0.01 to 24, with 1 being no change and 10 being equal to 0.1 in the magnitude of change.

Much thanks, i guess its time to take a stats class.

-jt

[yifangt]

Stats

You have three independent variables bacterial, gene and metabolism. One dependent variable Expression Folder Change. Better post your experiment design before a suggestion is given. Yifang

[jvanleuven]

Yes, three categorical independent variables. The first has 5 possible values, the second has 208 possible values, the third 164 possible values(summary(data) in R).

The same sample was treated in two different ways, then sequenced on a Hiseq machine. I have taken it though a number of processing steps to determine approximate coverage of transcripts (all denovo).

-jt

[yifangt]

I still seem not clear about your design. Correct me if I misunderstood anything:
Your data structure seems to be a matrix of col x row = 4 x 208, assuming all the combinations (not necessarily balanced, i.e. equal number of observation for each combination). It will not be surprising if there is missing data.

BACTERIUM GENE METABOLISM EXPR-FC
Bacteria1 Gene1 Metabolism 10
Bacteria1 Gene2 Metabolism 2.2
Bacteria2 Gene1 Metabolism 1.0
Bacteria2 Gene3 VitaminSynth 0.2
....
....

Now, the dependent variable is the "expression fold change" derived from Hiseq sequencing. I am expecting it is from the read counts of each gene, RPKM values, or other similar calculation, which needs be from at least two values of raw data (two different ways, as you said) for each "folder change". Following this,

I want to identify variables that are associated with the greatest fold change.

There will not be a simple answer to address this question. But, correlation analysis would be a first try, for which you may need to choose from options (Pearson's, Spearman's, partial correlations etc) depending on your variable properties.
As you have three independent variables, I would think of "multiple dimension reduction" strategy with linear multiple regression, principle component analysis or clustering analysis among others. There are many packages to do this job in R/Bioconductor.
About the normalization of your "folder change", it also depends on the method you choose. Simply you may not need to. Check with package limma and you will get more idea about analysis with "folder change" data which is actually a secondary data as a ratio of two raw numbers. That's another story.
Overall, this is a big complicate question, to me there is no simple easy answer. Good luck!

[jvanleuven]

Thank you yifangt. Hopefully this will get me started on the right path.