DSection Gene Expression Deconvolution Method
Description
The DSection algorithm performs gene expression deconvolution when priors on proportions are available, using a Markov Chain Monte Carlo approach Erkkila et al. (2010).
Usage
DSection(Y, p0, nSamples, groups = NULL, nBurnIn = 4 * nSamples, W0 = max(nrow(p0), 10), W_proposal = 100, samplep = TRUE, summarize = TRUE, verbose = FALSE)
Arguments
- Y
matrix of measurements from heterogeneous tissues. I is the number of probes/genes/etc., and J is the number of tissues. - 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.- groups
- <1-by-J> vector of treatment indices, so that unique(Treatment) = [1,2,...,C], where C is the number of treatments including control, i.e., "no treatment", if available.
- 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.
- nBurnIn
- Amount of burn-in. NOTE: keep nBurnIn > 0.
- nSamples
- Amount of sampling. NOTE: keep nSamples > 0.
- 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). - summarize
- logical indicating whether only average values should be returned -- and computed.
- verbose
- logical that indicates if verbose messages should be shown.
Value
A list with the following elements:
MCData results from the MCMC estimation. x_LS Standard-least square estimate. groups factor defining the groups of samples, if any was provided. call the call to DSection. parameters a list of some of the parameters used in the estimation. p0 initial prior on proportions.
Details
In CellMix, this method is registered with the key
'DSection', and is can be applied to gene
expression data via the function ged.
This function uses the RcppOctave package to run the original Matlab code in Octave. The documentation was extracted from the Matlab source file, that can be found in the CellMix package "scripts/DSection" subdirectory.
The Matlab code requires the Octave packages statistics and optim to run properly. These packages can be downloaded from Octave-forge:
http://sourceforge.net/projects/octave/files/Octave Forge Packages/Individual Package Releases/
and installed in Octave with:
pkg install '
or in R
install.extras('CellMix', 'octave:DSection')
References
http://informatics.systemsbiology.net/DSection/ http://www.cs.tut.fi/~erkkila2/software/dsection/index.html
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, 10)
dim(x)
## Features Samples Components
## 50 10 3
# extract true proportions
p <- coef(x)
# add noise to proportions
p0 <- scoef(abs(p + rmatrix(p, dist=rnorm, sd=0.15)))
# check how noisy this makes the proportion estimates
profplot(p, p0)
## Not run:
##D # requires octave package statistics to be installed (for gamfit)
##D res <- DSection(exprs(x), p0, nSamples=100, nBurnIn=1L)
##D profplot(p, res$MCData$p)
## End(Not run)