Question: statistics for differential expression: adjusted p-values<0.05 BUT negative B-odds?
gravatar for Christine Voellenkle
9.6 years ago by
Christine Voellenkle90 wrote:
Dear BioConductor mailing list! I am using the R-2.7.2, the limma package and its interface limmaGUI. I have a rather small number of slides (6) and probes spotted in 4 replicates (776x4=3104 spots). I perform background correction (normexp, cutoff=10), within (global loess) and between (Scale) array normalization. To obtain the statistics for differential expression I choose the "least squares" linear model fit and the calculation of Duplicate correlation, the adjust method for p-value is "BH". I get the following toptable: *Name* *P.Value* *logFC* *AveExpr* *t* *P.Value* *adj.P.Val* *B* hsa-miR-503 2.34E-06 1.945387328 8.136697759 5.884386792 2.34E-06 0.001343352 4.836590302 hsa-miR-921 3.46E-06 1.174433413 8.035845865 5.740377619 3.46E-06 0.001343352 4.471051142 hsa-miR-30c-2* 7.77E-06 3.117701794 8.27088038 5.445618379 7.77E-06 0.001550508 3.718681784 hsa-miR-198 7.99E-06 1.967637234 6.072614871 5.435302824 7.99E-06 0.001550508 3.692268736 miRPlus_42526 1.98E-05 3.697849577 9.172222106 5.105807836 1.98E-05 0.00307328 2.846724551 hsa-miR-665 2.55E-05 2.066123655 8.328358693 5.014730934 2.55E-05 0.003292112 2.612641297 hsa- miR-371-5p 6.28E-05 3.733502603 9.49054759 4.686906149 6.28E-05 0.006817575 1.770685081 hsa-miR-187* 7.03E-05 2.265561547 5.963832035 4.646081978 7.03E-05 0.006817575 1.666045526 hsa-miR-183* 9.64E-05 1.416875368 6.088668377 4.531017428 9.64E-05 0.008313829 1.371557327 hsa-miR-483-5p 0.000111163 1.789850517 6.054104314 4.479175715 0.000111163 0.008626216 1.239133367 hsa-miR-30b* 0.000123606 2.526268106 8.483475515 4.440467862 0.000123606 0.008719852 1.140379763 hsa-miR-620 0.00028076 1.84948372 8.448990163 4.139779212 0.00028076 0.018155808 0.377815095 miRPlus_17952 0.000337227 4.962342262 8.633804566 4.07218339 0.000337227 0.020129836 0.20777931 hsa-miR-675 0.000578682 1.991150452 5.45522505 3.871788436 0.000578682 0.028873358 -0.292416594 miRPlus_17869 0.000587344 1.280048919 8.085959229 3.866246126 0.000587344 0.028873358 -0.306158062 miRPlus_42793 0.000612763 1.315186076 6.393831483 3.850431131 0.000612763 0.028873358 -0.345339636 hsa-miR-193a-5p 0.000632535 1.418128062 7.751168126 3.838568191 0.000632535 0.028873358 -0.374700751 miRPlus_42487 0.00079838 5.437939056 10.51144667 3.751332648 0.00079838 0.033174042 -0.589807882 hsa-miR-637 0.000812251 1.586253029 5.38574605 3.744861076 0.000812251 0.033174042 -0.605707122 According to the book, p-values and B-statistics should rank genes in the same order. As possible treshhold for adjusted p-values <0.05, B-value of 0 expresses a 50:50 chance that its really differentially expressed, a negative B-value expresses a very very unlikey probability of differential expression. What makes me worry is that in my statistics I have low adj.p-value 0.03 together with negative B-values. How do I have to handle this discrepance? Is this a hint that something is wrong with my normalization? I performed validation by real-time PCR some time ago, at that time I considered only the p-value (and not adjusted .p or the B), using a treshhold of p<0.001. Now I checked these old results once again, to understand if positive B and adjusted p-values<0.05 in my case indiciated a high probability for modulated expression. It was true only for some mirRs (adjusted p-value < 0.05 an B positive) where modulation was confirmedby real time PCR. In contrast to other mirs (with adj.p > 0.05 and negative B) which showed modulation in real-time PCR. And yet other mir showed very good adjusted p and the B (adjp= 0.014, B=1.71)- but no modulation real-time. Are adjusted p-value and B-statistics too stringent or do I have to reconsider normlization and linear model fit? Do I expect too much from the Statistics? Grazie! Christine Dr. Christine Völlenkle, Ph.D. Research Laboratories-Molecular Cardiology I.R.C.C.S. Policlinico San Donato Via R. Morandi, 30 20097 S. Donato (MI) Italy Phone: +39 02 52774 683 (lab) +39 02 52774 533 (office) Fax: +39 02 52774 666 email: [[alternative HTML version deleted]]
ADD COMMENTlink modified 9.6 years ago by Paolo Innocenti320 • written 9.6 years ago by Christine Voellenkle90
gravatar for Paolo Innocenti
9.6 years ago by
Paolo Innocenti320 wrote:
> What makes me worry is that in my statistics I have low adj.p-value 0.03 > together with negative B-values. > How do I have to handle this discrepance? Is this a hint that something is > wrong with my normalization? In the limma userguide they also say: "The B-statistic is automatically adjusted for multiple testing by assuming that 1% of the genes, or some other percentage specified by the user in the call to eBayes(), are expected to be differentially expressed." and "The B-statistic probabilities depend on the same assumptions but require in addition a prior guess for the proportion of differentially expressed genes." So I think your problem is that the number of differentially expressed genes in your experiment is higher that the proportion eBayes assumes (proportion=0.01) Try to specify it in eBayes like this: fit <- eBayes(fit, proportion= <proportion of="" d.e.="" genes="">) and see if you get an improvement, and look at ?eBayes for more information. I never used limmaGUI though, so I don't know how to do it with this interface. I hope this helps. Best, paolo -- Paolo Innocenti Department of Animal Ecology, EBC Uppsala University Norbyv?gen 18D 75236 Uppsala, Sweden
ADD COMMENTlink written 9.6 years ago by Paolo Innocenti320
Please log in to add an answer.


Use of this site constitutes acceptance of our User Agreement and Privacy Policy.
Powered by Biostar version 2.2.0
Traffic: 193 users visited in the last hour