Hello,

I'm new to R and I have a little problem with data analysis. I was asked to create a correlation plot for 2 genes expressed in melanoma cells. I downloaded data from GEO (for 14 data sets, 2 types of similar microaarays), made Expression Sets, normalized with RMA, substracted data for 2 genes and compiled it into one matrix. To every sample I assigned two traits - cell type (normal, primary, metastasis and so on) and number of data set it was substracted from. Then I created simple plot to observe how my data looks like (without multiple sets normalization Spearman's correlation coefficient is above 0,5 with really low p-value). Now I would like to remove any differences between data sets - if I understood it correctly I should remove batch effect with e.g. ComBat. And here's my question - should I assume one batch equals one data set (or one data set contains more batches (differences in data collection dates and so on))? Is ComBat or SVA the best method for this particular case? And should I perform this normalization on whole data matrices (how?) and extract data for my 2 genes of interest?

I'm sorry if my post is a little chaotic but I'm still learning how to use R. I will be really greatful for your advice. 

Hello,

People assume of "what is batch" depending on their capabilities and/or precaution (it can be chip scan date*; dataset, microarray platform). Batch adjustments may significantly bias  group (phenotype) differences if your study design is unbalanced (groups are not evenly distributed across batches). From your description I deduce that you will have highly unbalanced design. There is many combination of steps you could take and none of them does  guarantee success.

It will be necessary for you to monitor your data (by using plots) after each important step.

On the assumption that you have unbalanced design and two platforms I would suggest one of the ways:

- normalize  each platform  (with RMA); use BrainArray mappings

- use Combat**  to merge data from both platforms (I assume that phenotypes on both platforms are more or less equal in therms of quantity) - here the platform is the batch.

or

- normalize  each platform (with RMA); use BrainArray mappings

- apply Combat for each platform separately (define "batch" as "data set" again be aware of phenotype composition of your "batches")

- perform scaling between platforms to merge data

or

You may try my suggestions without batch correction (eg. normalize and than scale). Sometimes the use of batch correction might be more harmful to data than not using it.

Cheers,

Pawel

*You can extract scan date from each .cel file

**Perform batch correction on the whole matrices. It is easer to get batch corrected matrix in Combat (that what you want for correlation analysis). SVA works  good in differential expression setting and is considered to be more "hard-core" to the data than Combat which makes it dangerous in the hands of inexperienced user.

Cześć Ewelina,

Please clarify:

Substracted means extracted?

14 datasets means 14 independent experiments (consisted of some number of samples, each)?

What kind of platforms they represent?

Best,

Pawel