Difference Between Mas5 And Rma Normalisation. Which Is More Appropriate When?

Question

[T. K.]

This question is rather general than specific. Hope, it's not too broad. I think a specific example is not required here.

I came up with this question a couple of month ago when analysing an Affymetrix microarray set of 16 cell lines. A colleague recommended me using MAS5 rather than RMA, because "it's more often used nowadays". Thus, I used MAS5 and not RMA.

However, I'm interested in a better reason for my choice (or mis-choice).

1. Where are the basic and important differences between MAS5 and RMA?

2. Are there (famous) examples, which show the advantages of the one over the other? Meaning, are there some general scenarios, where one should prefer the one over the other?

Thanks for your answers.

Answer 1

[Neilfws]

Ask any two bioinformaticians about microarray normalisation and you'll get 10 different answers :-)

A good summary of MAS5 versus RMA is provided in the article 'Summaries of Affymetrix GeneChip probe level data'. A slightly-less technical, but comprehensive review can be found in this PPT presentation. The essential differences between RMA and MAS5 are:

• MAS5 normalises each array independently and sequentially; RMA as the name suggests (robust multi-array) uses a multi-chip model
• MAS5 uses data from mismatch probes to calculate a "robust average", based on subtracting mismatch probe value from match probe value
• RMA does not use the mismatch probes, because their intensities are often higher than the match probes, making them unreliable as indicators of non-specific binding
• RMA values are in log2 units, MAS5 are not (so values are not directly comparable)

In the literature, you will always be able to find examples where people state that one method performed better than another; here's an article extolling the virtues of MAS5. The important thing to remember is that they observed the improvement precisely once, under a specific set of conditions - you can't generalise to all cases from one good result.

In general though, I disagree with your colleague: I'd say that RMA "is more often used nowadays."

I suggest searching the Web (Google for "rma mas5"), reading some of the literature (the journal Bioinformatics is a good source for these types of articles) and browsing the Bioconductor mailing list to get at least a feel for the discussion around different methods.

[User 59]

And MAS5 will return Present/Marginal/Absent flags on the data which can be used for filtering. I remember GeneSpring (now part of Agilent) saying that RMA/GCRMA produces less false positives on spike-in test data.

[T. K.]

Thank you for this summary. I've got plenty to read now :)

[Neilfws]

Right; the observation that spike-in probes behave comparably across chips in one experiment is the justification for the multi-chip model.

[T. K.]

@Daniel Thank you for this detail. The present/marginal/absent calls were the main point I needed during my work back then and probably another reason for my colleague recommending me MAS5.