Hi Ralittsa,

The normalization and normal transformation are very common procedures that are applied to microarray data in order to remove unwanted environmental bias from the experiment. This obviously changes the numeric appearance by scaling the numbers, but this methodology is proven to bring array experiments onto the same scale and make them comparable. Apart from the data transformation and normalization you might want to perform a PCA on your data (can be done in Gene spring).

Normal transformation

Microarray raw data are generally not ready to use for in data analysis and they require a few processing steps of which the normal transformation is an important step. The aim of a normal transformation is that data are transformed into a Gaussian distribution also called normal distribution. This enables the use of statistical methods that assume that data are normal distributed and it helps with the detection of technical glitches in the experiments. I guess there are some data visualization methods in Gene spring that would allow you display single array data and filter out the dodgy one if that is required.

Data normalization

Data normalization is a process that helps to remove environmental bias from experiments and brings all arrays on the same level so that they can be compared. There are different strategies for data normalization and the approach that was selected in this case has been tried and tested on protein array repeat experiments and the array experiments were analysed using mathematical correlation tests. The idea behind this experiment was that a repeat experiment should give near identical results and any deviation from the expected result in the raw data is caused by environmental effects. This data normalization study was performed on numerous occasions using batches of 24 arrays and different operators on different days.

The results of this experiments suggested that a data normalization procedure such as quantile, median or rank based normalization procedures that are beneficial for the analysis of expression arrays.

There are a few normalization methods around and quantile normalization if very effective (some would say harsh) . If you compare some repeat experiments before and after normalization. If their correlation (similarity ) after normalization improves you done the right thing. If it look worse than without normalization than try another method. 

If you have access to R or some decent data visualization software such Matlab or Tableau than you plot and test your data with those.

Kind regards,

Jens.

http://www.biomarker-discovery.com/data-analysis-overview

Dear Jens,

Many thanks for this extremely useful response! The lab that performed the analyses sent me the following information about the two approaches that they undertook:

"Both LIMMA and GENESPRING do the same statistical analysis. The big difference between the 2 analysis is that they use 2 different normalization methods at the beginning. For genespring we use 90% percentile shift while LIMMA uses a 75% quantile method. The Quantile method is slightly more permissive, hence why you get more microRNAs passing the FDR test."

I understand the basis of the 75 % quantile method, i.e. you divide by the signal value of the 75th percentile, but how does the 90 % percentile shift work in a way that makes it less permissive than the other? From what I understand from their information, they undertook either one of these analyses, but from your reply it would appear that they are used in combination? However, I got two different sets of results - one according to one method, the other according to the other (which is why I thought they were performed separately + based on the information I was sent). 

Thank you in advance for all this help!

Hi Ralitsa,

First of all it would be good to know how big the differences are.

Data analysis is a multi-step process and any step can influence the results.

In both analysis they use a different normalization method. Limma uses quantile normalization and gene spring used the percentile normalization. I doubt that this will make a huge difference (if it does than your signals are weak – maybe too weak to be of any use).

But there are other factors that influence the results are: number of dimensions / observations, outlier removal and RFU cut-off.

It is common to remove features with a high percentage of low RFU signals as they tend to produce high fold change hits that are not necessarily very trustworthy (high fold change in combination with high p-value).But there is no generally accepted rule where you draw the line - hence you can get differences in the result table.

In my personal opinion is the LIMMA package very optimistic and produces often many hits that can not be confirmed by other software packages.

So there are some hints:

a) look at the RFU values of your hits

b) make a list of  hits are in common in both data sets

c) have a look if you can trust the hits that are unique in each data set

If you have access to software that can produce boxplots or histograms you could work it out for yourself. 

Feel free to send me the RFU data and I can send you a Volcano plot with my hits.

Kind regards,

Jens

http://www.biomarker-discovery.com/

Dear Jens, 

Many thanks for your help! I have just asked the researcher who analysed the data (externally) to send me the normalised/cleaned RFU units. I will have a look at them as I have recently learned to use R and will try to perform a principal component analysis to look for patterns. Thank you for offering me your help - I would gladly send it to you if I get stuck! I will try to practice on my own first, but as I am still only learning the techniques, I may need some help. Is there a ready script that you would recommend? 

Best wishes, 

Ralitsa

Hi Ralitsa,

I am glad to hear that you want to take the business in your hands. I am not too much of an R specialist, but I believe there are some ready to use scripts for micro array analysis in Bioconductor. 

Kind regards,

Jens

http://www.biomarker-discovery.com/