Hi Michael, I'm posting this back to the Bioconductor list in case others search on E. coli, too. From my experience with that one experiment, the bimodality in the probe-level density plots (plotHist() in affycoretools is great!) does indicate the degree of degradation in your samples, especially the height of the first peak compared to the second peak. The one sample we knew was degraded but hybed anyway had a very small first peak and a very large second peak, and the samples that retained a lot of signal after GC background correction (see below) had larger first peaks and smaller second peaks. I also would be particularly worried if there was a lot of variation in your samples in the relative heights of the peaks. As for QC, using the fitPLM() function from the affyPLM package and then plotting the resulting weights (example code below) does help to show spatial effects on the arrays. For the pre-processing method, I would strongly suggest GCRMA, because the GC-based background correction removed almost ALL the signal from the one sample we knew was degraded; RT for prokaryotic samples uses random primers because there's no polyA tail, so all degraded fragments will get RT's and then labeled, and non-specific binding occurs more the higher the GC content, which is why the degraded sample had the HIGHEST raw signal values. For QC, I would first do GCRMA without normalization and do a boxplot on the resulting values to see which arrays still have signal and how much variation there is among arrays (example code below). I would probably exclude arrays that don't have much signal left, and if there's still a lot of variation in the distributions, I would think long and hard about using these values in the stat model, skipping the quantile normalization completely. As I said, I've only ever worked with this one experiment on E. coli, but we knew we were having problems with samples being degraded, and we decided to hyb one sample that was completely degraded, so I think my conclusions for this experiment were valid. I assume that they would be applicable to other E. coli experiments, but you'll have to decide that for yourself. HTH, Jenny > raw <- ReadAffy() > PLM.data <- fitPLM(raw) #this will fit RMA, but that's fine for spatial effects > image(PLM.data, which=1, type="weights") > image(PLM.data, which=2, type="weights") #etc. for all arrays > gcrma.nonorm <- gcrma(raw, normalize=FALSE) > boxplot(gcrma.nonorm) At 10:26 PM 8/21/2007, you wrote: >Hello Jenny, >I came across a discussion you posted (May, 2006) on the >Bioconductor forum about Affy E. Coli chips. I have recently been >looking at data from these chips for the first time, and observed >some of the same problems that you mention there (bimodality in >probe-level density plots, inapplicability of usual Affy QC tools), >and was wondering if you were able to arrive at anything like a >preferred analysis protocol, and if you might have some advice that >you could share. Information seems to be pretty sparse on these >arrays, so finding your submissions on the subject was very >encouraging. To be specific, I'm interested in any particular >methods you might have found most useful for array or RNA quality >assessment and a preferred normalization method (I observed a much >larger difference in RMA- vs. GCRMA-normalized data with these chips >than I have seen before). Also, can you tell me how you concluded >that the bimodality in the density plots was due to RNA degradation? >Thanks for any help, >Michael > >----- >Michael Slifker, MS >Biostatistics Facility >Fox Chase Cancer Center >333 Cottman Ave. >Philadelphia, PA 19111 >215-728-5345 > > Jenny Drnevich, Ph.D. Functional Genomics Bioinformatics Specialist W.M. Keck Center for Comparative and Functional Genomics Roy J. Carver Biotechnology Center University of Illinois, Urbana-Champaign 330 ERML 1201 W. Gregory Dr. Urbana, IL 61801 USA ph: 217-244-7355 fax: 217-265-5066 e-mail: drnevich at uiuc.edu