I have the utmost respect for Andrea Bacarelli, Vince Carey, and their colleagues, all of whom are extremely careful and methodical investigators. That said, I cannot help but wonder (based on my own data) whether characteristic differences in leukocyte populations, age, or chronic inflammation might be responsible for some of the observed population-level effect (given the large size of the cohort) in mixed blood cell populations. Then again, another possibility is that genuinely variable methylation regions are being pushed in one direction or another specifically as a result of COPD and its progression. Changes in blood cell populations themselves appear to be relatively small but replicable over the course of differentiation, with some notable exceptions. Consequently, I prefer to work with sorted cells for assessing DNA methylation markers whenever possible. When that is impossible, joint estimation of sample composition can be an attractive alternative. Nonetheless, flow sorting really ought not to break the bank in large studies like these (IMHO). Best, --t On Mon, Aug 20, 2012 at 1:29 PM, Brent Pedersen wrote: > On Mon, Aug 20, 2012 at 12:28 PM, Tim Triche, Jr. > wrote: > > > > > > On Mon, Aug 20, 2012 at 10:42 AM, Brent Pedersen > wrote: > >> > >> > >> > TCGA methylation data is background corrected and dye bias equalized > >> > (for > >> > the 450k samples, at least, and as batches are updated, 27k as well) > but > >> > no > >> > batch correction is done for the level 3 data. In the case of > >> > multi-batch > >> > tumors it is a good idea to run ComBat or (if you must) SVA to > >> > compensate. > >> > >> Sorry to hijack the thread, but, what is the reason to prefer ComBat > over > >> SVA? > > > > > > Because in practice, with calibration samples, it seems to work better. > > > > > >> > switching from 0.1% methylated to 99.9% methylated is probably a real > >> > effect. Switching from 1% to 3% across the board is probably > technical > >> > artifacts. > >> > >> I'm guessing this to be true only for tumor/normal comparisons or > >> "pure" samples. > > > > > > Yes, the former typically have distinct cancer-related (vs. > tissue-related) > > changes if any, and the latter are a bit like unicorn poop (never seen in > > the wild). > > > > > http://www.nature.com/nbt/journal/v30/n5/full/nbt.2203.html?WT.ec_id=NBT-201205#/methods > > > > > >> > >> What about peripheral blood where one may be measuring a signal from a > >> variety of cell types or tissues? > > > > > > Funny you should mention this particular task :-) > > > > > http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0041361 > > > > thanks for this reference, I hadn't seen it. Interesting to read that > after this study: > http://ajrccm.atsjournals.org/content/185/4/373.long > with tiny fold-changes that replicate across populations: > http://ajrccm.atsjournals.org/content/185/4/373/F3.large.jpg > > > > Given the difficulty of isolating gold standard reference populations by > > flow sorting, it's tough to benchmark the various transformations, but > what > > you gain in linearity you may lose in leverage. Since there isn't one > > particular transformation that simultaneously linearizes and stabilizes a > > proportion, > > > > > http://www.jstor.org/discover/10.2307/1269291?uid=2129&uid=2&uid=70&uid=4&sid=21101141681241 > > > > you have to pick your battles. In the case of compositional analysis, > 30+ > > years after Aitchison and Shen's seminal papers, it appears to remain > > unresolved. The ability to isolate a small number of highly purified > cells > > and perform targeted BS-seq on picogram quantities of DNA may put this to > > rest. > > > > > http://leg.est.ufpr.br/lib/exe/fetch.php/pessoais:abtmartins:thestatisticalanalysisofcompositionaldata.pdf > > > > However... joint analysis via DNA methylation and expression (array or > > RNAseq) is another matter, and there I have a candidate (in need of > > validation). > > > > I'll keep an eye out for that. > -b > > > > > I can't say that I'm entirely unhappy about you 'hijacking' this > thread... > > > > --t > > > > > -- *A model is a lie that helps you see the truth.* * * Howard Skipper [[alternative HTML version deleted]]