Loess Normalization And Scale Issues

Question

[Luca Beltrame]

Hello,

I'm normalizing a number of data files obtained from Agilent's miRNA platform. My simple procedure involves using GeneView files (which are processed by Feature Extraction before output) and then filter the miRNA probes by keeping only those with a positive flag in 60% of the samples.

EDIT: As the question wasn't clear, I need to add that the whole time, my data is either in a data.frame or a matrix.

The issue is that of course some "bad" probes for some samples will remain, and they might have 0 or negative intensities. My idea was to set those to NA.

I then proceeded to perform loess normalization (normalize.loess in affy): however, if I log2 the data beforehand (which is part of my workflow), even a single NA on a row will cause the whole row to be NA.

My question is then: should I get rid of those NAs altogether because normalize.loess is not fit to handle them? Of course I have no problems if I run the normalization in natural scale, but I'd be introducing biases and breaking assumptions.

Thanks in advance.

Answer 1

[Stefano Berri]

I don't know the nature of your data (if it is in an object or something) but usually there are two options The function has an option na.rm=FALSE as it happens for mean, sd. otherwise, if your data is in x you can do

y <- x[!is.na(x)]
y.loess <- normalize.loess(y)
x.loess <- x
x.loess[!is.na(x)] <- y.loess

Hope this will inspire you as I suspect your function does not allow you to tamber with data very much...

[Luca Beltrame]

Sorry, I forgot to mention that the data is either a matrix, or a data.frame.

[Stefano Berri]

Then this approach should work

Answer 2

[sysbiocoder]

You can also impute the values. There are several options. Please go through this tutorial http://r-statistics.co/Missing-Value-Treatment-With-R.html