superdelta: Super-delta sequence method for testing gene significance.
In fhlsjs/Super-delta: Super-delta-sequence method

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

The Super-delta sequence method is a combination of a) gene normalization procedure based on pairwise normalization; b) and a gene selection procedure based on a suitable test statistic. It must be applied to log-transformed, but not normalized gene expression level data. Only Student's two sample t-test is implemented in the current version.

1	superdelta(X, classlabel, test="t", side=c("two.sided", "less", "greater"), methods="robust", trim=0.2, baseline="auto", ...)

`X`	A data frame or matrix, with p rows corresponding to variables (hypotheses) and n columns to observations. In the case of gene expression data, rows correspond to genes and columns to mRNA samples. The data can be read using `read.table`.
`classlabel`	A vector of integers corresponding to observation (column) class labels. For k classes, the labels must be integers between 0 and k-1. We only implemented two-sample test as of 08/01/2015.
`test`	A character string specifying the statistic to be used to test the null hypothesis of no association between the variables and the class labels. If `test="t"`, the tests are based on Student's two-sample t-statistics (equal variances). If `test="f"`, the tests are based on F-statistics. If `test="pairt"`, the tests are based on paired t-statistics. We only implemented two-sample test as of 08/01/2015.
`side`	A character string specifying the type of rejection region. If `side="two.sided"`, two-tailed tests, the null hypothesis is rejected for large absolute values of the test statistic. If `side="more"`, one-tailed tests, the null hypothesis is rejected for large values of the test statistic. If `side="less"`, one-tailed tests, the null hypothesis is rejected for small values of the test statistic.
`methods`	A list of estimating methods to be used with the super-delta sequence method. The default value "robust" should be the best method for most applications. Other options include "mean", "median", and "mclust" (takes about 100 times more computing time). These advanced options are implemented for specialized situations only.
`trim`	A real number between 0 and 1. It controls the proportion of "extreme" test statistics to be exluded from the test statistic estimation step. This parameter is similiar in idea, but not identical to, the same parameter in function `mean()` for computing the robust sample mean.
`baseline`	See Section details.
`...`	Other advanced options. Currently only used for mclust method.

The super-delta sequence method is a combination of a) gene normalization procedure based on pairwise normalization; b) and a gene selection procedure based on a suitable test statistic. It must be applied to background corrected, log-transformed, but not normalized gene expression level data. Only Student's two sample t-test is implemented in the current version. Future plan includes paired t-test and ANOVA F-test. For more technical details of the super-delta sequence method please see the reference.

The trim parameter which controls how robust the super-delta sequence method is. In theory, a safe trim value should be no less than the proportion of true differentially expressed genes. The default value 0.2 should work for most practical applications for two reasons: a) it is not that common to have more than 20% genes selected as DEG; b) even if more than 20% of genes are DEG, chances are some of the "outliers" caused by pairing with the up-regulated genes get cancelled with those caused by the down-regulated genes; c) trimming is by nature a conservative operation, so some residual bias will not cause noticeable inflation of type I error. Please see the reference for more details.

The baseline paramter defines the set of baseline genes. If set to "all" or 0, every gene is normalized to every other gene. If set to a vector of integers, genes with these indices will be used as baseline genes, so every gene is normalized by these baseline genes only. This can save computation time and memory drastically for large data. If it is a positive integer, a random subset of genes of this size is created by sample(ngenes, baseline) and used as baseline genes. The default ("auto") is to use 1000 genes when the number of genes is larger than 1000 and all genes for smaller data.

If one method is specified (the default behavior), it produce a data frame with two components:

`teststat`	Vector (or matrix, if more than one method is specified) of test statistics.
`rawp`	Vector (or matrix) of raw (unadjusted) p-values.

If more than one method is specified, teststat and rawp are organized as data frames and put together as a list.

Xing Qiu, Yuhang Liu, Jinfeng Zhang

mean, t.test, p.adjust

## The Golub data.  On my computer, it takes about 3 seconds to run, YMMV.
## The result is a data.frame with two columns: teststat and rawp.

data(golub)
classlabel<-golub.cl
system.time(res1 <- superdelta(golub, classlabel))

## Golub data has 3,051 genes. You can use all genes as baseline genes,
## in which case it may take about 10 seconds to run.
## Not run: system.time(res2 <- superdelta(golub, classlabel, baseline="all"))

## use both robust and the mclust methods together with a lower-side
## test to find up-regulated genes.  A random set of 500 genes are used.
## The results are organized in a list of two data.frames instead. This
## command takes several minutes to run because Mclust is
## computationally very expensive.

## Not run: library(mclust); res3 <- superdelta(golub, classlabel,
methods=c("robust", "mclust"), side="lower", baseline=500)
## End(Not run)