Skip to main content ↑ ↓
Topic: Feature intensity drift correction suggestions (Read 5391 times) previous topic - next topic

Feature intensity drift correction suggestions

I've got around 100 samples that suffer from feature intensity drift. Even the first five QCs (so in a matter of 30 injections) show drift in feature intensity sequentially (so from QC1 to QC5 there's drift visible in the PCA).

Apart from intCor, which didn't really help, does anyone have any good suggestions on how to correct this please?

Appreciate any help at this point.

Thank you.

Re: Feature intensity drift correction suggestions

Reply #2 –
It might interesting to hear what has/hasn't worked for others in the forum.

I'll throw in my 2 cents.
Our lab has been experimenting with RUVIII (a variant of the 'PCA' method used by Wehrens). It can be found in the r package 'ruv' on cran.
RUVIII estimates the factors of unwanted variation using replicate (QC) samples and internal standards (or control metabolites).

So far it has been pretty effective at reducing the coefficient of variation of QC samples (not the sample QC samples used above, but rather a validation set).

Re: Feature intensity drift correction suggestions

Reply #3 –
Hi,
I would recommend you to use qc-svrc (the function is coded in matlab, if you're interested just send me an email to guira@uv.es). It's non-parametric (SVR) and its use is quite straightforward. Besides, it allows a fitting of the SVR parameters to the instrument performance by using the ε-insensitive loss parameter (for example, if you 'known' that a rsd~10% is acceptable in your instrument, then you can use that value to define the tolerance (as ε-insensitive loss parameter))
https://www.ncbi.nlm.nih.gov/pubmed/26462549
https://www.ncbi.nlm.nih.gov/pubmed/29852994

g.

Re: Feature intensity drift correction suggestions

Reply #4 –
I'm using linear models to estimate the drift for each feature (similar to Wehrens et al mentione by Jan above) on QC samples requiring that I have valid measurements from the QC samples for 3/4 of the injection index range within a measurement run.

I had also high hopes on the RUVIII approach, but it failed completely on the data set I tested it on. My best guess for this failure is that it needs internal standards to estimate the "unwanted variance", and the values I got from the internal standard were not representative for the "real" samples.

cheers, jo

Re: Feature intensity drift correction suggestions

Reply #5 –
Thanks for posting links to those articles, Guillermo.
The use of train-test (cross-validation) and a validate split for evaluation of batch correction is fantastic. Something that is often underappreciated.
I would like to try out qc-svrc, so I'll email you soon.

Jo, thanks for chiming in. That's an interesting point you make about RUVIII and something I'll keep an eye on.

It seems that linear models are being used by quite a few people, possibly with slight variations.
I ran some simulations using varying RSD% and number of QC samples. Overfitting is a possibility if QC numbers are low. Interestingly, including data from randomized samples reduces this chance. Perhaps including a weighting parameter in favor of QC samples might be prudent.

Cheers,
Corey