csaw (version 1.4.0)

clusterFDR: Compute the cluster-level FDR

Description

Compute the FDR across clusters based on the test-level FDR threshold

Usage

clusterFDR(ids, threshold)

Arguments

ids

an integer vector of cluster IDs for each significant test below threshold

threshold

a numeric scalar, specifying the FDR threshold used to define the significant tests

Value

A numeric scalar as the cluster-level FDR.

Details

This function computes an informal estimate of the cluster-level FDR, where each cluster is formed by aggregating only significant tests. In the context of ChIP-seq, each significant test refers to a DB window that is detected at a FDR below threshold. The idea is to obtain an error rate while reporting the precise coordinates of a DB subinterval in a complex region.

The cluster-level FDR is defined as the proportion of reported clusters that have no true positives. Simply using threshold is not appropriate, as the cluster- and window-level FDRs are not equivalent. This function also differs from the standard pipeline that is based on combineTests. Specifically, region definition in the standard pipeline must be independent of DB. Precise coordinates of the DB subinterval cannot be reported.

Users should note that the calculation of the cluster-level FDR here is not statistically rigorous. In particular, the observed number of false positive tests is estimated based on threshold and the total number of significant tests. This is not guaranteed to be an upper bound, especially when the observed window-level FDR is variable.

In conclusion, users should use the standard combineTests-based pipeline wherever possible. Clustering on significant windows should only be performed where the precise coordinates of the DB subinterval are important for interpretation.

See Also

mergeWindows, combineTests

Examples

# Setting up the windows and p-values.
windows <- GRanges("chrA", IRanges(1:1000, 1:1000))
test.p <- runif(1000)
test.p[c(1:10, 100:110, 220:240)] <- 0 # 3 significant subintervals.

# Defining significant windows.
threshold <- 0.05
is.sig <- p.adjust(test.p, method="BH") <= threshold

# Assuming that we only cluster significant windows.
merged <- mergeWindows(windows[is.sig], tol=0)
clusterFDR(merged$id, threshold)