Become an expert in R — Interactive courses, Cheat Sheets, certificates and more!
Get Started for Free

predFC

Predictive log-fold changes


Description

Computes estimated coefficients for a NB glm in such a way that the log-fold-changes are shrunk towards zero.

Usage

## S3 method for class 'DGEList'
predFC(y, design, prior.count=0.125, offset=NULL, dispersion=NULL, 
       weights=NULL, ...)
## S3 method for class 'SummarizedExperiment'
predFC(y, design, prior.count=0.125, offset=NULL, dispersion=NULL, 
       weights=NULL, ...)
## Default S3 method:
predFC(y, design, prior.count=0.125, offset=NULL, dispersion=0, 
       weights=NULL, ...)

Arguments

y

a matrix of counts, or a DGEList object, or a SummarizedExperiment object

design

the design matrix for the experiment

prior.count

the average prior count to be added to each observation. Larger values produce more shrinkage.

offset

numeric vector or matrix giving the offset in the log-linear model predictor, as for glmFit. Usually equal to log library sizes.

dispersion

numeric vector of negative binomial dispersions.

weights

optional numeric matrix giving observation weights

...

other arguments are passed to glmFit.

Details

This function computes predictive log-fold changes (pfc) for a NB GLM. The pfc are posterior Bayesian estimators of the true log-fold-changes. They are predictive of values that might be replicated in a future experiment.

Specifically, the function adds a small prior count to each observation before fitting the GLM (see addPriorCount for details). The actual prior count that is added is proportion to the library size. This has the effect that any log-fold-change that was zero prior to augmentation remains zero and non-zero log-fold-changes are shrunk towards zero.

The prior counts can be viewed as equivalent to a prior belief that the log-fold changes are small, and the output can be viewed as posterior log-fold-changes from this Bayesian viewpoint. The output coefficients are called predictive log fold-changes because, depending on the prior, they may be a better prediction of the true log fold-changes than the raw estimates.

Log-fold changes for genes with low counts are shrunk more than those for genes with high counts. In particular, infinite log-fold-changes arising from zero counts are avoided. The exact degree to which this is done depends on the negative binomail dispersion.

Value

Numeric matrix of (shrunk) linear model coefficients on the log2 scale.

Author(s)

Belinda Phipson and Gordon Smyth

References

Phipson, B. (2013). Empirical Bayes modelling of expression profiles and their associations. PhD Thesis. University of Melbourne, Australia. http://repository.unimelb.edu.au/10187/17614

See Also

Examples

# generate counts for a two group experiment with n=2 in each group and 100 genes
disp <- 0.1
y <- matrix(rnbinom(400,size=1/disp,mu=4), nrow=100, ncol=4)
y <- DGEList(y, group=c(1,1,2,2))
design <- model.matrix(~group, data=y$samples)

#estimate the predictive log fold changes
predlfc <- predFC(y, design, dispersion=disp, prior.count=1)
logfc <- predFC(y,design,dispersion=disp, prior.count=0)
logfc.truncated <- pmax(pmin(logfc,100),-100)

#plot predFC's vs logFC's
plot(predlfc[,2], logfc.truncated[,2],
     xlab="Predictive log fold changes", ylab="Raw log fold changes")
abline(a=0,b=1)

edgeR

Empirical Analysis of Digital Gene Expression Data in R

v3.32.1
GPL (>=2)
Authors
Yunshun Chen, Aaron TL Lun, Davis J McCarthy, Matthew E Ritchie, Belinda Phipson, Yifang Hu, Xiaobei Zhou, Mark D Robinson, Gordon K Smyth
Initial release
2021-01-14

We don't support your browser anymore

Please choose more modern alternatives, such as Google Chrome or Mozilla Firefox.