michaelisaiahlove at gmail.com
Wed Mar 13 09:45:43 CET 2013
Thanks for the comments.
On 03/13/13 08:59, Ryan C. Thompson wrote:
> I noticed the addition of the DESeq2 package a few days ago and had a
> look at the new additions. Overall, it looks like an excellent package
> (and it runs a lot faster than DESeq 1, too). I have a few questions
> for clarification of exactly what methods DESeq2 is using.
> Specifically, I notice that the fold change shrinkage is performed in
> nbinomWaldTest but not in nbinomChisqTest. Is this just for reasons of
> backward-compatibility of results, or is the Chi-squared test
> logically incompatible with shrunken GLM coefficients?
The latter was our reason. With shrunken coefficients, the differences
in deviances are no longer distributed as a chi-square. For example,
with simulated null data (non-intercept betas equal to zero), the
differences in deviances will pile up near zero and the resulting
p-values will not be uniformly distributed.
> Secondly, is there any plan to extend the Wald test to testing
> contrasts of multiple coefficients or testing multiple
> coefficients/contrasts at once in an ANOVA-like test?
Yes, we are looking into a convenient interface for this.
> -Ryan Thompson
> On 03/12/2013 01:51 PM, Wolfgang Huber wrote:
>> Dear DESeq users,
>> Mike Love, Simon Anders and I have been updating the DESeq package.
>> This resulted in the package DESeq2, which is available from the
>> development branch, and scheduled for the next release:
>> For several release cycles, the original package (DESeq) will be
>> maintained at its current functionality, in order to not disrupt the
>> workflows of DESeq users. For new projects, we recommend using
>> DESeq2. Major innovations are:
>> * Base class: SummarizedExperiment is used as the superclass for
>> storing the data, rather than eSet. This allows closer integration
>> with upstream workflows involving GRanges and summarizeOverlaps, and
>> facilitates downstream analyses of the genomic regions of interest.
>> * Simplified workflow: the wrapper function DESeq() performs all
>> steps for a differential expression analysis. The individual steps
>> are of course also accessible.
>> * More powerful statistics: incorporation of prior distributions into
>> the estimation of dispersions and fold changes (empirical-Bayes
>> shrinkage). The dispersion shrinkage improves power compared to the
>> old DESeq. The fold changes shrinkage help moderate the otherwise
>> large spread in log fold changes for genes with low counts, while it
>> has negligible effect on genes with high counts; it may be
>> particularly useful for visualisation, clustering, classification,
>> ordination (PCA, MDS), similar to the variance-stabilizing
>> transformation in the old DESeq. A Wald test for significance is
>> provided as the default inference method, with the chi-squared test
>> of the previous version is also available. A manuscript is in
>> * Normalization: it is possible to provide a matrix of sample- *and*
>> gene-specific normalization factors, which allows the use of
>> normalisation factors from Bioconductor packages such as cqn and EDASeq.
>> Examples of usage are provided in the vignette, and more details are
>> available in the manual pages (specifically, the DESeq function and
>> estimateDispersions function).
>> Enjoy -
>> Mike, Simon, Wolfgang.
>> Bioconductor mailing list
>> Bioconductor at r-project.org
>> Search the archives:
> Bioconductor mailing list
> Bioconductor at r-project.org
> Search the archives:
More information about the Bioconductor