roast: Rotation Gene Set Tests

In hdeberg/limma: Linear Models for Microarray Data

Description

Rotation gene set testing for linear models.

Usage

## Default S3 method:
roast(y, index = NULL, design = NULL, contrast = ncol(design), geneid = NULL,
      set.statistic = "mean", gene.weights = NULL, var.prior = NULL, df.prior = NULL,
      nrot = 1999, approx.zscore = TRUE, legacy = FALSE, ...)
## Default S3 method:
mroast(y, index = NULL, design = NULL, contrast = ncol(design), geneid = NULL,
       set.statistic = "mean", gene.weights = NULL, var.prior = NULL, df.prior = NULL,
       nrot = 1999, approx.zscore = TRUE, legacy = FALSE, adjust.method = "BH",
       midp = TRUE, sort = "directional", ...)
## Default S3 method:
fry(y, index = NULL, design = NULL, contrast = ncol(design), geneid = NULL,
      gene.weights = NULL, standardize = "posterior.sd", sort = "directional", ...)

Arguments

y	numeric matrix giving log-expression or log-ratio values for a series of microarrays, or any object that can coerced to a matrix including ExpressionSet, MAList, EList or PLMSet objects. Rows correspond to probes and columns to samples. NA or infinite values are not allowed. If either var.prior or df.prior are NULL, then y should contain values for all genes on the arrays. If both prior parameters are given, then only y values for the test set are required.
index	index vector specifying which rows (probes) of y are in the test set. Can be a vector of integer indices, or a logical vector of length nrow(y), or a vector of gene IDs corresponding to entries in geneid. Alternatively it can be a data.frame with the first column containing the index vector and the second column containing directional gene contribution weights. For mroast or fry, index is a list of index vectors or a list of data.frames.
design	design matrix
contrast	contrast for which the test is required. Can be an integer specifying a column of design, or the name of a column of design, or a numeric contrast vector of length equal to the number of columns of design.
geneid	gene identifiers corresponding to the rows of y. Can be either a vector of length nrow(y) or the name of the column of y$genes containing the gene identifiers. Defaults to rownames(y).
set.statistic	summary set statistic. Possibilities are "mean","floormean","mean50" or "msq".
gene.weights	numeric vector of directional (positive or negative) contribution weights specifying the size and direction of the contribution of each probe to the gene set statistics. For mroast or fry, this vector must have length equal to nrow(y). For roast, can be of length nrow(y) or of length equal to the number of genes in the test set.
var.prior	prior value for residual variances. If not provided, this is estimated from all the data using squeezeVar.
df.prior	prior degrees of freedom for residual variances. If not provided, this is estimated using squeezeVar.
nrot	number of rotations used to compute the p-values. Low values like 999 are suitable for testing but higher values such as 9999 or more are recommended for publication purposes.
approx.zscore	logical, if TRUE then a fast approximation is used to convert t-statistics into z-scores prior to computing set statistics. If FALSE, z-scores will be exact.
legacy	logical. See Note below for usage.
adjust.method	method used to adjust the p-values for multiple testing. See p.adjust for possible values.
midp	logical, should mid-p-values be used in instead of ordinary p-values when adjusting for multiple testing?
sort	character, whether to sort output table by directional p-value ("directional"), non-directional p-value ("mixed"), or not at all ("none").
standardize	how to standardize for unequal probewise variances. Possibilities are "residual.sd", "posterior.sd" or "none".
...	any argument that would be suitable for lmFit or eBayes can be included.

Details

These functions implement rotation gene set tests proposed by Wu et al (2010). They perform self-contained gene set tests in the sense defined by Goeman and Buhlmann (2007). For competitive gene set tests, see camera. For a gene set enrichment analysis (GSEA) style analysis using a database of gene sets, see romer.

roast and mroast test whether any of the genes in the set are differentially expressed. They can be used for any microarray experiment that can be represented by a linear model. The design matrix for the experiment is specified as for the lmFit function, and the contrast of interest is specified as for the contrasts.fit function. This allows users to focus on differential expression for any coefficient or contrast in a linear model. If contrast is not specified, then the last coefficient in the linear model will be tested.

The argument index is often made using ids2indices but does not have to be. Each set to be tested is represented by a vector of row numbers or a vector of gene IDs. Gene IDs should correspond to either the rownames of y or the entries of geneid.

All three functions support directional contribution gene weights, which can be entered either through the gene.weights argument or via index. Directional gene weights allow each gene to be flagged as to its direction and magnitude of change based on prior experimentation. A typical use is to make the gene.weights 1 or -1 depending on whether the gene is up or down-regulated in the pathway under consideration. Probes with directional weights of opposite signs are expected to have expression changes in opposite directions. Gene with larger gene weights in absolute size will have more weight in the set statistic calculation.

Gene weights can be either genome-wide or set-specific. Genome-wide weights can be entered via the gene.weights argument. Set specific weights can be input by including the gene weights as part of the set's entry in index. If any of the components of index are data.frames, then the second column will be assumed to be gene contribution weights for that set. All three functions (roast, mroast and fry) support set-specific gene contribution weights as part of an index data.frame.

Set-specific directional gene weights are used to represent expression signatures assembled from previous experiments, from gene annotation or from prior hypotheses. In the output from roast, mroast or fry, a significant "Up" p-value means that the differential expression results found in y are positively correlated with the expression signature coded by the gene weights. Conversely, a significant "Down" p-value means that the differential expression log-fold-changes are negatively correlated with the expression signature.

Note that the contribution weights set by gene.weights are different in nature and purpose to the precision weights set by the weights argument of lmFit. gene.weights control the contribution of each gene to the formation of the gene set statistics and are directional, i.e., can be positive or negative. weights indicate the precision of the expression measurements and should be positive. The weights are used to construct genewise test statistics whereas gene.weights are used to combine the genewise test statistics.

The arguments df.prior and var.prior have the same meaning as in the output of the eBayes function. If these arguments are not supplied, then they are estimated exactly as is done by eBayes.

The gene set statistics "mean", "floormean", "mean50" and msq are defined by Wu et al (2010). The different gene set statistics have different sensitivities when only some of the genes in a set are differentially expressed. If set.statistic="mean" then the set will be statistically significantly only when the majority of the genes are differentially expressed. "floormean" and "mean50" will detect as few as 25% differentially expressed in a set. "msq" is sensitive to even smaller proportions of differentially expressed genes, if the effects are reasonably large. Overall, the "msq" statistic gives the best power for rejecting the null hypothesis of no differentially expressed genes, but the significance can be driven by a small number of genes. In many genomic applications it is appropriate to limit results to gene sets for which most of the genes response in a concordance direction, so the relatively conservative "mean" statistic is the default choice.

The output gives p-values three possible alternative hypotheses, "Up" to test whether the genes in the set tend to be up-regulated, with positive t-statistics, "Down" to test whether the genes in the set tend to be down-regulated, with negative t-statistics, and "Mixed" to test whether the genes in the set tend to be differentially expressed, without regard for direction.

roast estimates p-values by simulation, specifically by random rotations of the orthogonalized residuals (Langsrud, 2005), so p-values will vary slightly from run to run. The p-value is computed as (b+1)/(nrot+1) where b is the number of rotations giving a more extreme statistic than that observed (Phipson and Smyth, 2010). This means that the smallest possible mixed or two-sided p-values are 1/(nrot+1). The function uses a symmetry argument to double the effective number of rotations for the one-sided tests, so the smallest possible "Up" or "Down" p-value is 1/(2*nrot+1).

The number of rotations nrot can (and should) be increased tTo get more precise p-values from roast or mroast, The default nrot is set fairly low to facilitate quick testing and experimentation but the smallest possible two-sided p-value is 1/(nrot+1). To get definitive p-values for publication, at least nrot=9999 or higher is recommended.

mroast does roast tests for multiple sets, including adjustment for multiple testing. By default, mroast reports ordinary p-values but uses mid-p-values (Routledge, 1994) at the multiple testing stage. Mid-p-values are probably a good choice when using false discovery rates (adjust.method="BH") but not when controlling the family-wise type I error rate (adjust.method="holm").

To improve the performance of the gene set statistics, roast and mroast transform the genewise moderated t-statistics to normality using zscoreT. By default, an approximate closed-form transformation is used (approx.zscore=TRUE), which is very much faster than the exact transformation and performs just as well. In Bioconductor 2.10, the transformation used has been changed from Hill's (1970) approximation to Bailey's (1980) formula because the latter is faster and gives more even accuracy; see zscoreT for more details.

fry is a fast alternative designed to approximate what mroast with set.stat="mean" would give for a very large (infinite) number of rotations. In the special case that df.prior is very large and set.statistic="mean", fry gives the same directional p-values that mroast would give if an infinite number of rotations could be performed. In other circumstances, when genes have different variances, fry uses a standardization strategy to approximate the mroast results. Using fry is recommended when performing tests for a large number of sets because it is fast and because it returns higher resolution p-values that are not limited by the number of rotations performed. Note, the close approximation of fry to mroast is only for the directional p-values. The fry mixed p-values are computed by a different method and will not necessarily be very close to those from mroast.

Value

roast produces an object of class "Roast". This consists of a list with the following components:

p.value	data.frame with columns Active.Prop and P.Value, giving the proportion of genes in the set contributing materially to significance and estimated p-values, respectively. Rows correspond to the alternative hypotheses Down, Up, UpOrDown (two-sided) and Mixed.
var.prior	prior value for residual variances.
df.prior	prior degrees of freedom for residual variances.

mroast produces a data.frame with a row for each set and the following columns:

NGenes	number of genes in set
PropDown	proportion of genes in set with z < -sqrt(2)
PropUp	proportion of genes in set with z > sqrt(2)
Direction	direction of change, "Up" or "Down"
PValue	two-sided directional p-value
FDR	two-sided directional false discovery rate
PValue.Mixed	non-directional p-value
FDR.Mixed	non-directional false discovery rate

fry produces the same output format as mroast but without the columns PropDown and ProbUp.

Note

For Bioconductor 3.10, roast and mroast have been revised to use much less memory by conducting the rotations in chunks and to be about twice as fast by updating the normalizing transformation used when approx.zscore=TRUE. For a limited time, users wishing to reproduce Bioconductor 3.9 results exactly can set legacy=TRUE to turn these revisions off.

approx.score=TRUE become the default in Bioconductor 3.0 (October 2014).

The default set statistic was changed from "msq" to "mean" in Bioconductor 2.7 (October 2010).

Author(s)

Gordon Smyth and Di Wu

References

Goeman, JJ, and Buhlmann, P (2007). Analyzing gene expression data in terms of gene sets: methodological issues. Bioinformatics 23, 980-987.

Langsrud, O (2005). Rotation tests. Statistics and Computing 15, 53-60.

Phipson B, and Smyth GK (2010). Permutation P-values should never be zero: calculating exact P-values when permutations are randomly drawn. Statistical Applications in Genetics and Molecular Biology, Volume 9, Article 39. http://www.statsci.org/smyth/pubs/PermPValuesPreprint.pdf

Routledge, RD (1994). Practicing safe statistics with the mid-p. Canadian Journal of Statistics 22, 103-110.

Wu, D, Lim, E, Francois Vaillant, F, Asselin-Labat, M-L, Visvader, JE, and Smyth, GK (2010). ROAST: rotation gene set tests for complex microarray experiments. Bioinformatics 26, 2176-2182. http://bioinformatics.oxfordjournals.org/content/26/17/2176

See Also

See 10.GeneSetTests for a description of other functions used for gene set testing.

Examples

y <- matrix(rnorm(100*4,sd=0.3),100,4)
design <- cbind(Intercept=1,Group=c(0,0,1,1))

# First set of 5 genes are all up-regulated
index1 <- 1:5
y[index1,3:4] <- y[index1,3:4]+3
roast(y,index1,design,contrast=2)

# Second set of 5 genes contains none that are DE
index2 <- 6:10
mroast(y,list(set1=index1,set2=index2),design,contrast=2)
fry(y,list(set1=index1,set2=index2),design,contrast=2)

# Third set of 6 genes contains three down-regulated genes and three up-regulated genes
index3 <- 11:16
y[index3[1:3],3:4] <- y[index3[1:3],3:4]-3
y[index3[4:6],3:4] <- y[index3[4:6],3:4]+3

# Without gene weights
# Mixed p-value is significant for set3 but not the directional p-values
mroast(y,list(set1=index1,set2=index2,set3=index3),design,contrast=2)
fry(y,list(set1=index1,set2=index2,set3=index3),design,contrast=2)

# With gene weights
# Set3 is significantly up (i.e., positively correlated with the weights)
index3 <- data.frame(Gene=11:16,Weight=c(-1,-1,-1,1,1,1))
mroast(y,list(set1=index1,set2=index2,set3=index3),design,contrast=2)
fry(y,list(set1=index1,set2=index2,set3=index3),design,contrast=2)

hdeberg/limma documentation built on Dec. 20, 2021, 3:43 p.m.