fitGG: Fit GaGa hierarchical model

Description Usage Arguments Details Value Author(s) References See Also Examples

View source: R/fitGG.r

Description

fitGG fits GaGa/MiGaGa hierarchical models, either via a fully Bayesian approach or via maximum likelihood.

fitNN fits a normal-normal hierarchical model (wrapper to call emfit in package EBarrays with the LNNMV model). fitNNSingleHyp is the same as fitNN but only considers the pattern that all groups are different from each other.

adjustfitNN corrects a small sample-size bias in the fitNN estimation procedure.

Usage

1
2
3
4
5
6
7
8
fitGG(x, groups, patterns, equalcv = TRUE, nclust = 1, method =
 "quickEM", B, priorpar, parini, trace = TRUE)

fitNN(x, groups, patterns, B=20, trace=TRUE)

fitNNSingleHyp(x, groups, B=10, trace=TRUE)

adjustfitNN(fit, pitrue, B=5, nsim=3, mc.cores=1)

Arguments

x

ExpressionSet, exprSet, data frame or matrix containing the gene expression measurements used to fit the model. For fitNN, x should be in log-scale (in contrast to the input to emfit). fitGG accepts raw and log-scale (as long as all log-measurements remain positive). By default we recommend using log-scale to reduce the influence of outliers.

groups

If x is of type ExpressionSet or exprSet, groups should be the name of the column in pData(x) with the groups that one wishes to compare. If x is a matrix or a data frame, groups should be a vector indicating to which group each column in x corresponds to.

patterns

Matrix indicating which groups are put together under each pattern, i.e. the hypotheses to consider for each gene. colnames(patterns) must match the group levels specified in groups. Defaults to two hypotheses: null hypothesis of all groups being equal and full alternative of all groups being different. The function buildPatterns can be used to construct a matrix with all possible patterns.

equalcv

equalcv==TRUE fits model assuming constant CV across groups. equalcv==FALSE compares cv as well as mean expression levels between groups

nclust

Number of clusters in the MiGaGa model. nclust corresponds to the GaGa model.

method

method=='MH' fits a fully Bayesian model via Metropolis-Hastings posterior sampling. method=='Gibbs' does the same using Gibbs sampling. method=='SA' uses Simulated Annealing to find the posterior mode. method=='EM' finds maximum-likelihood estimates via the expectation-maximization algorithm, but this is currently only implemented for nclust>1. method=='quickEM' is a quicker implementation that only performs 2 optimization steps (see details).

B

Number of iterations to fit the model. For method=='MH' and method=='Gibbs', B is the number of MCMC iterations (defaults to 1000). For method=='SA', B is the number of iterations in the Simulated Annealing scheme (defaults to 200). For method=='EM', B is the maximum number of iterations (defaults to 20).

priorpar

List with prior parameter values. It must have components a.alpha0,b.alpha0,a.nu,b.nu,a.balpha,b.balpha,a.nualpha,b.nualpha,p.probclus and p.probpat. If missing they are set to non-informative values that are usually reasonable for RMA and GCRMA normalized data.

parini

list with components a0, nu, balpha, nualpha, probclus and probpat indicating the starting values for the hyper-parameters. If not specified, a method of moments estimate is used.

trace

For trace==TRUE the progress of the model fitting routine is printed.

fit

nnfit object, as returned by fitNN

pitrue

Grid of true pi values for which to evaluate the MoM estimation bias. See details.

nsim

Number of datasets to simulate for each pitrue value

mc.cores

If package multicore is available, mc.cores specifies the number of cores to be used for parallel computing.

Details

For GaGa/MiGaGa models, an approximation is used to sample faster from the posterior distribution of the gamma shape parameters and to compute the normalization constants (needed to evaluate the likelihood). These approximations are implemented in rcgamma and mcgamma.

The cooling scheme in method=='SA' uses a temperature equal to 1/log(1+i), where i is the iteration number.

The EM implementation in method=='quickEM' is a quick EM algorithm that usually delivers hyper-parameter estimates very similar to those obtained via the slower method=='EM'. Additionally, the GaGa model inference has been seen to be robust to moderate changes in the hyper-parameter estimates in most datasets.

fitNN is a wrapper to emfit in package EBarrays with the LNNMV model. This procedure estimates hyper-parameters using the method of moments (MoM), which typically results in over-estimating the proportion of differentially expressed genes, which we denote by pi1. adjustfitNN corrects this bias by repeatedly simulating from the prior predictive of a normal-normal model. Simulations are performed for a grid of pi1 values, so that the expected bias can be evaluated for each of them. The bias is then modeled as a smooth function of pi1 using function gam from package mgcv. Finally, the value of pi1 is bias-adjusted and the posterior probabilities are recomputed using the updated pi1 value.

Value

fitGG returns an object of class gagafit, with components

parest

Hyper-parameter estimates. Only returned if method=='EBayes', for method=='Bayes' one must call the function parest after fitGG

mcmc

Object of class mcmc with posterior draws for hyper-parameters. Only returned if method=='Bayes'.

lhood

For method=='Bayes' it is the log-likelihood evaluated at each MCMC iteration. For method=='EBayes' it is the log-likelihood evaluated at the maximum.

nclust

Same as input argument.

patterns

Same as input argument, converted to object of class gagahyp.

fitNN returns an analogous object of class nnfit. The component nn.fit is the object returned by emfit.

Author(s)

David Rossell

References

Rossell D. (2009) GaGa: a Parsimonious and Flexible Model for Differential Expression Analysis. Annals of Applied Statistics, 3, 1035-1051.

Yuan, M. and Kendziorski, C. (2006). A unified approach for simultaneous gene clustering and differential expression identification. Biometrics 62(4): 1089-1098.

See Also

parest to estimate hyper-parameters and compute posterior probabilities after a GaGa or MiGaGa fit. findgenes to find differentially expressed genes. classpred to predict the group that a new sample belongs to.

Examples

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
library(gaga)
set.seed(10)
n <- 100; m <- c(6,6)
a0 <- 25.5; nu <- 0.109
balpha <- 1.183; nualpha <- 1683
probpat <- c(.95,.05)
xsim <- simGG(n,m,p.de=probpat[2],a0,nu,balpha,nualpha,equalcv=TRUE)
x <- exprs(xsim)

#Frequentist fit: EM algorithm to obtain MLE
groups <- pData(xsim)$group[c(-6,-12)]
patterns <- matrix(c(0,0,0,1),2,2)
colnames(patterns) <- c('group 1','group 2')
gg1 <- fitGG(x[,c(-6,-12)],groups,patterns=patterns,method='EM',trace=FALSE)  
gg1 <- parest(gg1,x=x[,c(-6,-12)],groups)
gg1

gaga documentation built on Nov. 8, 2020, 5:49 p.m.