Timeseries loop design analysis using Limma or Maanova?
2
0
Entering edit mode
Pete ▴ 70
@pete-486
Last seen 8.8 years ago
Hello all, I have been asked to analyse a set of timecourse data with an unusual incomplete loop design. This is the design of this type I have looked at and I'm not entirely sure how to treat it. The initial (and fairly easy) question asked of the data is, what are the differences between the mutant and the control animals at each timepoint? The second question is how the mutant changes across the timeseries. The authors wish to use a bayesian timeseries clustering algorithmn to analyse this, but this requires a standardised measure for the mutant at each timepoint. I am unsure quite how to achieve this second point and welcome any suggestions or references that may help. Is this something I could do in Limma or MAanova? The data are from spotted, two-colour, oligo arrays. There are 6 timepoints. At each timepoint, tissue samples from 3 individual mutant animals are compared to a control pool of WT animals at the same timepoint, with dye swaps. In addition each control pool has then been compared in a dye swap to the next timepoint control pool. See diagram below (if it comes out correctly!) or the table further below where a1 a2 a3 represent any 3 individual animals. a1t1 a2t1 a3t1 a1t2 a2t2 a3t2 etc............ \\ || // \\ || // Control t1 ========= Control t2 ==== etc............... or SLIDE CY3 CY5 1 a1t1 control t1 2 control t1 a1t1 3 a2t1 control t1 4 control t1 a2t1 5 a3t1 control t1 6 control t1 a3t1 7 a1t2 control t2 8 control t2 a1t2 9 a2t2 control t2 10 control t2 a2t2 11 a3t2 control t2 12 control t2 a3t2 13 a1t3 control t3 14 control t3 a1t3 15 a2t3 control t3 16 control t3 a2t3 17 a3t3 control t3 18 control t3 a3t3 19 a1t4 control t4 20 control t4 a1t4 21 a2t4 control t4 22 control t4 a2t4 23 a3t4 control t4 24 control t4 a3t4 25 a1t5 control t5 26 control t5 a1t5 27 a2t5 control t5 28 control t5 a2t5 29 a3t5 control t5 30 control t5 a3t5 31 a1t6 control t6 32 control t6 a1t6 33 a2t6 control t6 34 control t6 a2t6 35 a3t6 control t6 36 control t6 a3t6 37 control t1 control t2 38 control t2 control t1 39 control t2 control t3 40 control t3 control t2 41 control t3 control t4 42 control t4 control t3 43 control t4 control t5 44 control t5 control t4 45 control t5 control t6 46 control t6 control t5 Many thanks Pete
0
Entering edit mode
Jenny Drnevich ★ 2.2k
@jenny-drnevich-382
Last seen 8.8 years ago
At 10:07 AM 2/13/2006, Pete wrote: >Hello all, > >I have been asked to analyse a set of timecourse data with an unusual >incomplete loop design. This is the design of this type I have looked at >and I'm not entirely sure how to treat it. > >The initial (and fairly easy) question asked of the data is, what are the >differences between the mutant and the control animals at each timepoint? I am interested in how you are going to analyze the differences between mutants and controls at each time point given that there is no replication of the control animals (only 1 control pool). I just advised a researcher against this kind of experimental design because I could not think offhand of a way to analyze it statistically. If there is a statistically valid method, I would like to know about it. >The second question is how the mutant changes across the timeseries. The >authors >wish to use a bayesian timeseries clustering algorithmn to analyse this, but >this requires a standardised measure for the mutant at each timepoint. How are you going to implement this bayesian timeseries clustering? My interpretation of clustering algorithms in general is that they should not be used to determine which genes are "differentially" expressed, but rather one should first use a statistical model to determine differential expression, then only cluster the genes that show a significant difference somewhere along the time series to find groups of genes that show a similar expression pattern. My approach to this situation would be something along the lines of a single-channel analysis using a mixed model with array + dye + treatment + time + treatment*time, and then cluster genes that showed a significant time effect, using the lsmeans for each mutant*timepoint group. The lack of replication of the controls may cause this not to work... Cheers, Jenny >I am unsure quite how to achieve this second point and welcome any >suggestions or references that may help. Is this something I could do in >Limma or MAanova? > > >The data are from spotted, two-colour, oligo arrays. There are 6 timepoints. >At each timepoint, tissue samples from 3 individual mutant animals are >compared to a control pool of WT animals at the same timepoint, with dye >swaps. In addition each control pool has then been compared in a dye swap to >the next timepoint control pool. See diagram below (if it comes out >correctly!) or the table further below where a1 a2 a3 represent any 3 >individual animals. > > > >a1t1 a2t1 a3t1 a1t2 a2t2 a3t2 etc............ > \\ || // \\ || // > Control t1 ========= Control t2 ==== etc............... > >or > >SLIDE CY3 CY5 >1 a1t1 control t1 >2 control t1 a1t1 >3 a2t1 control t1 >4 control t1 a2t1 >5 a3t1 control t1 >6 control t1 a3t1 >7 a1t2 control t2 >8 control t2 a1t2 >9 a2t2 control t2 >10 control t2 a2t2 >11 a3t2 control t2 >12 control t2 a3t2 >13 a1t3 control t3 >14 control t3 a1t3 >15 a2t3 control t3 >16 control t3 a2t3 >17 a3t3 control t3 >18 control t3 a3t3 >19 a1t4 control t4 >20 control t4 a1t4 >21 a2t4 control t4 >22 control t4 a2t4 >23 a3t4 control t4 >24 control t4 a3t4 >25 a1t5 control t5 >26 control t5 a1t5 >27 a2t5 control t5 >28 control t5 a2t5 >29 a3t5 control t5 >30 control t5 a3t5 >31 a1t6 control t6 >32 control t6 a1t6 >33 a2t6 control t6 >34 control t6 a2t6 >35 a3t6 control t6 >36 control t6 a3t6 >37 control t1 control t2 >38 control t2 control t1 >39 control t2 control t3 >40 control t3 control t2 >41 control t3 control t4 >42 control t4 control t3 >43 control t4 control t5 >44 control t5 control t4 >45 control t5 control t6 >46 control t6 control t5 > > >Many thanks > >Pete > >_______________________________________________ >Bioconductor mailing list >Bioconductor at stat.math.ethz.ch >https://stat.ethz.ch/mailman/listinfo/bioconductor 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
0
Entering edit mode
Pete ▴ 70
@pete-486
Last seen 8.8 years ago
Thanks for your response, >>Hello all, >> >>I have been asked to analyse a set of timecourse data with an unusual >>incomplete loop design. This is the design of this type I have looked at >>and I'm not entirely sure how to treat it. >> >>The initial (and fairly easy) question asked of the data is, what are the >>differences between the mutant and the control animals at each timepoint? > > I am interested in how you are going to analyze the differences between > mutants and controls at each time point given that there is no replication > of the control animals (only 1 control pool). I just advised a researcher > against this kind of experimental design because I could not think offhand > of a way to analyze it statistically. If there is a statistically valid > method, I would like to know about it. > I'm not quite sure I understand your point here? I was going to treat this as a simple dye swap experiment, ignoring time and comparing mutant to WT. Is this not a statistically valid approach? There are 3 independ mutant samples compared in dyeswaps to the WT pool. I understand that there is no biological replicate for the WT pool, however it is technically replicated at the dyeswap level and cDNA synthesis level. The biological variation of the WT population is not of immediate interest in this case, hence a pool was used. Individual mutant samples were used instead of a pool, because only a limited number of mutants were available. > >>The second question is how the mutant changes across the timeseries. The >>authors >>wish to use a bayesian timeseries clustering algorithmn to analyse this, >>but >>this requires a standardised measure for the mutant at each timepoint. > > How are you going to implement this bayesian timeseries clustering? My > interpretation of clustering algorithms in general is that they should not > be used to determine which genes are "differentially" expressed, but > rather > one should first use a statistical model to determine differential > expression, then only cluster the genes that show a significant difference > somewhere along the time series to find groups of genes that show a > similar > expression pattern. My approach to this situation would be something along > the lines of a single-channel analysis using a mixed model with array + > dye > + treatment + time + treatment*time, and then cluster genes that showed a > significant time effect, using the lsmeans for each mutant*timepoint > group. > The lack of replication of the controls may cause this not to work... > > Cheers, > Jenny I agree with your statement about clustering, and prehaps I didn't word my question very clearly. The timeseires clustering will indeed be performed on genes selected as differential with respect to time. In the past I have used the MAANOVA package to select these differential genes, however in that particular case, the samples were all compared back to a single reference sample rather than multiple references that are then compared to eachother in an incomplete loop. The issue I am concerned with is how to both, select genes that have a time effect, and how/what to use as a standardised expression level for these genes so that it can then be used in the clustering. Cheers Pete > > > >>I am unsure quite how to achieve this second point and welcome any >>suggestions or references that may help. Is this something I could do in >>Limma or MAanova? >> >> >>The data are from spotted, two-colour, oligo arrays. There are 6 >>timepoints. >>At each timepoint, tissue samples from 3 individual mutant animals are >>compared to a control pool of WT animals at the same timepoint, with dye >>swaps. In addition each control pool has then been compared in a dye swap >>to >>the next timepoint control pool. See diagram below (if it comes out >>correctly!) or the table further below where a1 a2 a3 represent any 3 >>individual animals. >> >> >> >>a1t1 a2t1 a3t1 a1t2 a2t2 a3t2 etc............ >> \\ || // \\ || // >> Control t1 ========= Control t2 ==== etc............... >> >>or >> >>SLIDE CY3 CY5 >>1 a1t1 control t1 >>2 control t1 a1t1 >>3 a2t1 control t1 >>4 control t1 a2t1 >>5 a3t1 control t1 >>6 control t1 a3t1 >>7 a1t2 control t2 >>8 control t2 a1t2 >>9 a2t2 control t2 >>10 control t2 a2t2 >>11 a3t2 control t2 >>12 control t2 a3t2 >>13 a1t3 control t3 >>14 control t3 a1t3 >>15 a2t3 control t3 >>16 control t3 a2t3 >>17 a3t3 control t3 >>18 control t3 a3t3 >>19 a1t4 control t4 >>20 control t4 a1t4 >>21 a2t4 control t4 >>22 control t4 a2t4 >>23 a3t4 control t4 >>24 control t4 a3t4 >>25 a1t5 control t5 >>26 control t5 a1t5 >>27 a2t5 control t5 >>28 control t5 a2t5 >>29 a3t5 control t5 >>30 control t5 a3t5 >>31 a1t6 control t6 >>32 control t6 a1t6 >>33 a2t6 control t6 >>34 control t6 a2t6 >>35 a3t6 control t6 >>36 control t6 a3t6 >>37 control t1 control t2 >>38 control t2 control t1 >>39 control t2 control t3 >>40 control t3 control t2 >>41 control t3 control t4 >>42 control t4 control t3 >>43 control t4 control t5 >>44 control t5 control t4 >>45 control t5 control t6 >>46 control t6 control t5 >> >> >>Many thanks >> >>Pete >> >>_______________________________________________ >>Bioconductor mailing list >>Bioconductor at stat.math.ethz.ch >>https://stat.ethz.ch/mailman/listinfo/bioconductor > > 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 > > _______________________________________________ > Bioconductor mailing list > Bioconductor at stat.math.ethz.ch > https://stat.ethz.ch/mailman/listinfo/bioconductor >
0
Entering edit mode
Hi Pete, Pete wrote: > I'm not quite sure I understand your point here? I was going to treat this > as a simple dye swap experiment, ignoring time and comparing mutant to WT. > Is this not a statistically valid approach? There are 3 independ mutant > samples compared in dyeswaps to the WT pool. I understand that there is no > biological replicate for the WT pool, however it is technically replicated > at the dyeswap level and cDNA synthesis level. The biological variation of > the WT population is not of immediate interest in this case, hence a pool > was used. Individual mutant samples were used instead of a pool, because > only a limited number of mutants were available. You can certainly do something like this, but there are some caveats. First, by comparing WT to mutant and ignoring time you are essentially looking at a main effect that might not be of much interest (hence why would you make the effort to do a time series?). Usually a more interesting question is to look for genes that are differentially expressed between mutant and WT at particular times, which I assume is why Jenny said you have no replication. Second, when you compare biological replicates to technical replicates you are underestimating the true variability of the WT samples, which may result in apparent significance where there may have been none had biological replicates been used for WT samples as well. This is usually only a problem when you try to validate the results (using new biologically replicated samples), if there are many genes that fail to validate. Since the validation step is usually much slower and laborious, decreasing the number of false positives in the microarray step is often worth the time and effort. Best, Jim -- James W. MacDonald Affymetrix and cDNA Microarray Core University of Michigan Cancer Center 1500 E. Medical Center Drive 7410 CCGC Ann Arbor MI 48109 734-647-5623