Hi
I have studied some papers on meta-analysis of microarray data. Some of authors recommend to normalize two-color data by lowess and one-color data with RMA algorithm. As we know lowess normalization method removes intensity dependent bias while there are another types of bias like spatial, dye and scale bias. I think we should compare data-sets with minimal systematic variation and biases, and therefore it is not important which normalization method is applied. I'd like to know different ideas.
Best,
Samane
Yes, IF you have data-sets with minimal systematic variation, THEN there would no need to normalize at all.
lowess adjusts the intensity profiles *within* an array, not between arrays. This is an entirely different task than RMA is supposed to to, namely a normalization *between* arrays. You can compare RMA to quantile-normalization or to vsn, for instance, but not to lowess.
Thanks for your reply. No I mean that we pre-process data to minimize systematic variations. It is not correct to normalize all two color platforms with lowess and one color with RMA. Sometime another normalization methods minimize biases in better way, for example vsn for one of my two color data and combination of a within-array and between-array normalization method for another data with the same platform. I believe it is not correct to select a normalization method for all data with similar platform because of different types of biases.
I can import series matrix (pre-processed data in GEO) into R and I am almost sure that data producer has tried to minimize variations with suitable pre-processing methods. However, in some meta analysis papers have been mentioned that you should follow a consistent normalization method in similar platforms!!!
Using different data series from GEO is a very dangerous thing. It is likely that you will have a confounding of treatments/conditions and series (like condition A was measured (only) in series X and condition B (only) in series Y - contrasting A vs. B will always be confounded by any systematic difference between X vs, Y as well, and there is no chance to estimate or remove this confounding effect).
You often can not sensibly remove differences between series, because the sample material (e.g. RNA) could be of considerably different quality (e.g. one is from microdissected samples, another is from cell cultures), amplified and labelled with different protocols, and measured with probes against different regions of the gene.
It is clear that normalization is done just to "remove" such differences, but in these cases it will introduce additional problems. It will be ok as long as series are compared with material of similar quality, processed in the same way, hybridized on similar arrays, and that have all sample types hybridised in all series (->confounding). Otherwise the results from such "meta analyses2 should be taken with extreme care (some worthwile insights may still be possible, but there is a great danger of over-interpretation).
I would further not use the pre-processed data but always use the raw data (cel-files, spot-files, gpr-files etc).
Many thanks.........You are totally right about negative points of meta-analysis. In regard to your reasons, normalization may not remove biases among different series but we are able to minimize these systematic variations among arrays (at least replicated arrays) within one set using normalization. I believed doing normalization or another pre-processing methods like background correction or filtering is better than not doing. However your ideas has made me think more!!!!!!
I am happy if I could make you think more :)
The key point, I think, is the confounding. If sample types/groups/conditions are confounded with batches/series, normalization can be even harmful or deleterious. You may find many huge and crystal clear effects (gene regulations) between conditions, but you have no means to find out if these effects are due to the (tenchnological) differences between batches/series or due to (biological) differences the conditions.
If there is no confounding then it may just happen that the noise is increased and one won't see any clear effects. In such cases it might be better to include the batch/series ID as additional predictor, if possible including the interactions with the conditions (since the effects don't need to be of the same size for all batches/series). Using hierarchical models may help to retain enough "degrees of freedom" for the analysis of the interesting effects.
Thank you for your time and consideration. I will try to consider all of these nice points in meta-analysis. But integration of microarray data with different platforms is still challenging for me.
It is challanging (not just for you), and there is possibly no really good solution. So think carfully, communicate carefully what you did, don't overinterpret the results, and make sure what you did is the best you could have done based on the available data, your possibilities and knowledge. That's science ;)
wow......a nice definition for "science"
I am thinking of using a criterion like mean of variance among replicated arrays or RT-PCR results to evaluate pre-processing methods in data sets of my interest. I will try to do a precise evaluation in each step. Thank you again
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Meta-Analysis