Partial Gene Expression Deconvolution with DSection
Description
Estimates cell/tissue cell/tissue-specific expression
signatures, given proportion priors using the MCMC
approach from Erkkila et al. (2010), implemented
in Matlab and wrapped in CellMix by the function
DSection.
The S3 method csTopTable for DSection fits
computes nominal p-values (i.e. unadjusted) of
differential expression between cell type or group of
samples within each cell type, for deconvolution results
from the DSection algorithm.
Usage
gedAlgorithm.DSection(..., maxIter = 500L, p0, data = NULL, nBurnIn = 4 * maxIter,
W0 = max(100, nbasis(seed)), W_proposal = 100, samplep = TRUE)
S3 (DSection)
`csTopTable`(x, coef = 1L, decreasing = TRUE, ...)
Arguments
- data
- variable (e.g., factor) that defines groups of samples. Cell-specific signatures will be computed for each group.
- maxIter
- number of sampling, not including burn-in sampling.
- ...
- extra arguments passed to
DSection. - coef
- specifies the reference cell type or group of samples, for which differential expression is computed.
- p0
matrix of prior predictions on cell type proportions. T is the number of cell types, and columns in p0must be positive and add up to one.- nBurnIn
- Amount of burn-in. NOTE: keep nBurnIn > 0.
- W0
- Prior prediction weight, i.e., degree of
confidence, on
p0. Defines the peakedness of Dirichlet density around p0. NOTE: keep W0 >= T. - W_proposal
- Transition kernel weight, defines the peakedness of Dirichlet density around p*, the old value. The higher W_proposal is, the smaller the proposal steps around p* are.
- samplep
- logical value, indicating whether to
sample from the posterior for cell type proportions
(
TRUE) or not (FALSE). SUGGESTED USE: sample from the posterior (samplep = 1). - x
- data object, typically returned by
ged. - decreasing
- logical that indicates the feature ordering, based on their p-values or FDRs.
Details
The DSection algorithm requires the Octave
statistics package to be installed (more
specifically the function gamfit). Unfortunately
some very recent changes in R or Octave
prevent RcppOctave from running DSection's
Matlab code (due to some error linked with Octave
io package). It is therefore temporary unavailable.
References
Erkkila T, Lehmusvaara S, Ruusuvuori P, Visakorpi T,
Shmulevich I and Lahdesmaki H (2010). "Probabilistic
analysis of gene expression measurements from
heterogeneous tissues." _Bioinformatics (Oxford,
England)_, *26*(20), pp. 2571-7. ISSN 1367-4811,
Examples
# random global expression
x <- rmix(3, 50, 20)
# extract true proportions
p <- coef(x)
# add noise
p0 <- scoef(abs(p + rmatrix(p, dist=rnorm, sd=0.15)))
# check how noisy this makes the proportion estimates
profplot(p, p0)
# deconvolve using DSection
## Not run:
##D # requires octave package statistics to be installed (for gamfit)
##D res <- ged(x, p0, 'DSection', maxIter=10, seed=12345)
##D head(basis(res))
##D # proportions are updated
##D !identical(coef(res), p0)
##D # check how better they are
##D profplot(x, res)
##D
##D \dontshow{
##D stopifnot( !identical(coef(res), p0) )
##D stopifnot( nmf.equal(res, ged(x, p0, 'DSection', maxIter=10, seed=12345)) )
##D }
## End(Not run)