check_resid: Check the residuals of a DESeq model

In dswatson/biowrapr: Convenient wrappers for functions in bioinformatics analysis pipelines

Description

This function evaluates the impact of different expression filters on the residuals of a fitted DESeqDataSet.

Usage

check_resid(dds, filter = c(1, 1))

Arguments

dds	A DESeqDataSet object that has been fit with a negative binomial GLM.
filter	Numeric vector of length two specifying the filter criterion. Each probe must have at least filter[1] log2-counts per million in at least filter[2] libraries to pass the expression threshold.

Details

The statistical tests upon which qmod is based presume that residuals for a fitted model are approximately normally distributed. This is not generally true of negative binomal GLMs, the family of models used by DESeq. The non-normality of residuals is especially pronounced for low count probes, which are by default not filtered out until after modeling in the DESeq pipeline. (See results for more details.) To run qmod on DESeqDataSet objects, it is necessary to filter out underexpressed probes and apply a variance stabilizing transformation. We recommend applying the lightest possible expression filter, although there is no precise algorithm for determining what this should be.

As a rule of thumb, the limma authors advise setting filter[1] to 10 / (L / 1,000,000), where L = the minimum library size for a given count matrix; and setting filter[2] to the number of replicates in the largest group. These are broad guidelines, however, not strict rules.

Examples

library(DESeq2)
dds <- makeExampleDEESeqDataSet()
dds <- DESeq(dds)
check_resid(dds)

dswatson/biowrapr documentation built on May 15, 2019, 4:52 p.m.