SigCheck-package: Check a gene signature's classification performance against random signatures, permuted data, and known signatures.

Description

While gene signatures are frequently used to classify data (e.g. predict prognosis of cancer patients), it is not always clear how optimal or meaningful they are (cf David Venet, Jacques E. Dumont, Vincent Detours' paper "Most Random Gene Expression Signatures Are Significantly Associated with Breast Cancer Outcome"). Based on suggestions in that paper, SigCheck accepts a data set (as an ExpressionSet) and a gene signature, and compares its classification performance (using the MLInterfaces package) against a) random gene signatures of the same length; b) permuted data; and c) known, unrelated gene signatures.

Arguments

Details

Package:

SigCheck

Type:

Package

Version:

1.0

Date:

2014-06-26

License:

Artistic-2.0

SigCheck provides a high-level function, sigCheck, that runs all the core functions in turn. The four core functions enable 1) a genes signature's baseline classification performance to be established (sigCheckClassifier), 2) compares performance against signatures composed of random genes (sigCheckRandom), 3) compares performance against known, and mostly unrelated, gene signatures (sigCheckKnown), and 4) compares performance against randomly permuted data (sigCheckPermuted).

At a minimum, SigCheck requires a data set (as an ExpressionSet) and a signature (a subset of features in the ExpressionSet). It uses the MLearn funciton formt he MLInterfaces package to build a classifier (using link{smvI} by default) and measure its performance against validation samples in the ExpressionSet; if no validation samples are specified, it uses leave-one-out (LOO) cross-validation to build multiple classifiers, each predicting one sample.

Output of each check includes the distribution of random performance scores (percentage of validation samples correctly classified) and the ranking of the passed signature in this distribution. A simple p-value calculation based on this rank is also returned.

References

Venet, David, Jacques E. Dumont, and Vincent Detours. "Most random gene expression signatures are significantly associated with breast cancer outcome." PLoS Computational Biology 7.10 (2011): e1002240.