So my understanding is that if there are no technical replicates one could do a single channel analysis in LIMMA using the duplicate correlation command to indicate the 2 samples on the same array. This would be equivalent to having a random effect for array - hence allowing the simplicity of the single channel analysis with a statistically appropriate means of handling the within spot correlation. --Naomi At 09:48 PM 6/24/2005, Gordon K Smyth wrote: >On Sat, June 25, 2005 12:36 am, Wolfgang Huber said: > >> Basically, you're saying that if the arrays are very high quality, you can > >> get away with an inefficient analysis. > > > > Gordon, I did not say that, it sounds stupid, please do not misquote > > people. > >Actually I didn't quote you at all. The word "Basically" in this context >is a signal that I am >interpreting your comments and their consequences rather than quoting >you. You can disagree with >my interpretation or can argue that it is mistaken, as you do below, but >being misquoted is quite >a different thing! :) > > >> Naomi is refering to what I call the "intraspot" correlation, see for > >> example the intraspotCorrelation() function in the limma package, and it > >> is critically important. The correlation isn't a bad thing, nor is it > >> restricted to poor quality arrays. Rather it means that contrasts > >> estimated within a spot are highly accurate. > > > > I agree that contrasts estimated from within one array are more > > accurate than those from different arrays. > >And in order to combine these two types of contrasts efficiency in an >analysis, one needs to >quantify the difference in accuracy. Hence the need to estimate the >intraspot correlation. > > > Note that when I said > > "treat a two-color array like two single-color arrays", this was in > > the paragraph on how to normalize, not on differential expression. But > > apparently this still triggered off a few people ... > >Part of the trouble is that you continued on in the next paragraph to >consider differential >expression, and you seemed (to me at least) to be implying that the same >conclusions continue to >apply with only one caveat. Thanks for the clarification. > >As you know, I personally prefer to take advantage of the two-colour >technology even at the >normalisation stage, but that's another matter. > > > Two aspects were raised by Claus' question that started this thread: > > how to normalize these data, and how to identify differentially > > expressed genes. My experience is that multi-channel normalization > > methods like vsn (or quantiles for that matter) work well for sets of > > mass-produced two-color arrays. Then, it is still better to look at > > contrasts within arrays. But it is at least possible (even if less > > accurate / precise) to look at contrasts across arrays by directly > > comparing the intensities, rather than always having to go through a > > chain of log-ratios. > >Claus' asked what is specific to Agilent. As I understand it, all your >comments here apply to any >type of two-colour array. Did you intend to say something specific about >Agilent arrays or am I >still mis-understanding what you mean? > > >> Why not do it properly and get the full benefit of the high > >> quality arrays? My experience is that high quality > >> Agilent arrays can beat affy for accuracy if treated properly. > > > > Agreed. Do you think it's because of the two colors or of the longer > > (and hence more specific) probes ? > >Well, Affy actually has more nucleotides per gene than Agilent when one >takes into account the >multiple probes per probe set. I don't want to speculate too much on the >reasons, but the fact >that Agilent can reliably lay down 80mers rather than 25mers strongly >suggests that the deposition >process is more accurate. The two colours are certainly >important. Calculations in our lab >suggest that one typically loses around 70% of information in a two colour >experiment by going >from direct to indirect comparisons, and 80-90% when going to single >channel comparisons across >different arrays without taking the intraspot correlations into >account. So Agilent may be well >behind Affy if not treated optimally. > >Gordon > > > Best wishes > > Wolfgang > > > > <quote who="Gordon Smyth"> > >> Wolfgang, > >> > >> Naomi is refering to what I call the "intraspot" correlation, see for > >> example the intraspotCorrelation() function in the limma package, and it > >> is > >> critically important. The correlation isn't a bad thing, nor is it > >> restricted to poor quality arrays. Rather it means that contrasts > >> estimated > >> within a spot are highly accurate. It is what makes the two- colour > >> technology intrinsically more accurate than one channel technology, other > >> things being equal. See http://www.statsci.org/smyth/pubs/ISI2005-116.pdf > >> for some discussion. > >> > >> Basically, you're saying that if the arrays are very high quality, you can > >> get away with an inefficient analysis. Why not do it properly and get the > >> full benefit of the high quality arrays? My experience is that high > >> quality > >> Agilent arrays can beat affy for accuracy if treated properly. > >> > >> Gordon > >> > >>>Date: Thu, 23 Jun 2005 15:29:38 +0100 (BST) > >>>From: "Wolfgang Huber" <huber at="" ebi.ac.uk=""> > >>>Subject: Re: [BioC] Agilent Arrays > >>>To: "Naomi Altman" <naomi at="" stat.psu.edu=""> > >>>Cc: bioconductor at stat.math.ethz.ch > >>> > >>>Hi Naomi, > >>> > >>>and why is that important? Also, what is the within gene correlation > >>>between green foreground of array 1 and green foreground of array 2? > >>> > >>>Bw > >>> Wolfgang > >>> > >>><quote who="Naomi Altman"> > >>> > I am working with Agilent arrays on which we have spotted many > >>> replicates > >>> > of the control spots. > >>> > The within gene correlation between red and green forground is about > >>> 0.8 > >>> > for the unnormalized data - i.e. pretty high! > >>> > > >>> > --Naomi > >>> > > >>> > At 03:23 AM 6/23/2005, Wolfgang Huber wrote: > >>> >>Hi Claus, > >>> >> > >>> >>for the normalization of arrays where the spotting etc. variability > >>> >>between chips is not strong, you can treat the data from m two-colour > >>> >>arrays as if it were 2*m single colour ones, and use methods like > >>> >>"quantiles" or "vsn". > >>> >> > >>> >>Note that for almost all genes, the hybridization is not limited by > >>> the > >>> >>amount of probe DNA, hence the competition between red and gree target > >>> is > >>> >>negligible for almost all genes (execept possibly the most highly > >>> >>expressed ones). This justifies treating a two-color array like two > >>> >>single-color arrays. > >>> >> > >>> >>Only later when you consider the contrasts of interest for finding > >>> >>differentially expressed genes, you want to make sure that these are > >>> not > >>> >>confounded with dye. > >>> >> > >>> >>PS, I think your question is very directly Bioconductor related! > >>> >> > >>> >>Best wishes > >>> >> Wolfgang > >>> >> > >>> >> > >>> >><quote who="Claus Mayer"> > >>> >> > Dear all! > >>> >> > > >>> >> > Apologies for asking a question which is not directly Bioconductor > >>> >> > related: After some experience with spotted 2-channel arrays and > >>> >> > Affydata, I am currently analysing my first data set based on > >>> Agilent > >>> >> > arrays. I know that packages like marray or limma have facilities > >>> to > >>> >> > read these data and that they can be normalised and analysed like > >>> any > >>> >> > other 2-colour-arrays. On the other hand the printing technology of > >>> >> > these arrays (using inkjet-printing of 60mer oligos) is closer in > >>> >> spirit > >>> >> > to Affy, if I understand this correctly. This seems to show in the > >>> >> data > >>> >> > as well. For example the strongest correlations I found in the > >>> single > >>> >> > channel (log-)intensities was not between the two channels observed > >>> on > >>> >> > the same slide (like with spotted arrays), but between the two > >>> >> channels > >>> >> > (differently dyed on different arrays in a loop design) that > >>> contained > >>> >> > the same sample (which is quite reassuring). This made me wonder > >>> >> whether > >>> >> > (once dye and array effects have been removed by some normalisation > >>> >> > method) with Agilent arrays one might really use single channel > >>> >> > intensities as measures of gene expression instead of reducing them > >>> to > >>> >> > the log-ratio only as is usually done for two-channel data. > >>> >> > > >>> >> > This would have consequences on the way these arrays should be > >>> >> > normalised (rather by a multichip method than individually) and > >>> also > >>> >> > allow more flexibility in the design of experiments. > >>> >> > > >>> >> > As I said before this is my first Agilent data set, so I would be > >>> >> > interested to hear opinions of others with more experience. Before > >>> I > >>> >> > start to re-invent the wheel here, I?d be also interested to know > >>> >> > whether any of you is aware of tools, software, papers, etc? > >>> dealing > >>> >> > with the analysis of Agilent array data specifically (rather than > >>> just > >>> >> > applying standard methods for 2-coloured cDNA -arrays). > >>> >> > > >>> >> > Any help/comments appreciated > >>> >> > > >>> >> > Claus > >>> >> > > >>> >> > -- > >>> >> > > >>> >> > >>> > ******************************************************************** *************** > >>> >> > Claus-D. Mayer | http://www.bioss.ac.uk > >>> >> > Biomathematics & Statistics Scotland | email: claus at bioss.ac.uk > >>> >> > Rowett Research Institute | Telephone: +44 (0) 1224 > >>> 716652 > >>> >> > Aberdeen AB21 9SB, Scotland, UK. | Fax: +44 (0) 1224 715349 > >>> >> > > >>> >> > _______________________________________________ > >>> >> > Bioconductor mailing list > >>> >> > Bioconductor at stat.math.ethz.ch > >>> >> > https://stat.ethz.ch/mailman/listinfo/bioconductor > >>> >> > > >>> >> > > >>> >> > >>> >> > >>> >>------------------------------------- > >>> >>Wolfgang Huber > >>> >>European Bioinformatics Institute > >>> >>European Molecular Biology Laboratory > >>> >>Cambridge CB10 1SD > >>> >>England > >>> >>Phone: +44 1223 494642 > >>> >>Http: www.ebi.ac.uk/huber > >>> >> > >>> >>_______________________________________________ > >>> >>Bioconductor mailing list > >>> >>Bioconductor at stat.math.ethz.ch > >>> >>https://stat.ethz.ch/mailman/listinfo/bioconductor > >>> > > >>> > Naomi S. Altman 814-865-3791 (voice) > >>> > Associate Professor > >>> > Bioinformatics Consulting Center > >>> > Dept. of Statistics 814-863-7114 (fax) > >>> > Penn State University 814-865-1348 > >>> (Statistics) > >>> > University Park, PA 16802-2111 Naomi S. Altman 814-865-3791 (voice) Associate Professor Bioinformatics Consulting Center Dept. of Statistics 814-863-7114 (fax) Penn State University 814-865-1348 (Statistics) University Park, PA 16802-2111