vsn (version 3.40.0)

normalize.AffyBatch.vsn: Wrapper for vsn to be used as a normalization method with expresso

Description

Wrapper for vsn2 to be used as a normalization method with the expresso function of the package affy. The expresso function is deprecated, consider using justvsn instead. The normalize.AffyBatch.vsn can still be useful on its own, as it provides some additional control of the normalization process (fitting on subsets, alternate transform parameters).

Usage

normalize.AffyBatch.vsn( abatch, reference, strata = NULL, subsample = if (nrow(exprs(abatch))>30000L) 30000L else 0L, subset, log2scale = TRUE, log2asymp=FALSE, ...)

Arguments

abatch
An object of type AffyBatch.
reference
Optional, a 'vsn' object from a previous fit. If this argument is specified, the data in 'x' are normalized "towards" an existing set of reference arrays whose parameters are stored in the object 'reference'. If this argument is not specified, then the data in 'x' are normalized "among themselves". See vsn2 for details.
strata
The 'strata' functionality is not supported, the parameter is ignored.
subsample
Is passed on to vsn2.
subset
This allows the specification of a subset of expression measurements to be used for the vsn fit. The transformation with the parameters of this fit is then, however, applied to the whole dataset. This is useful for excluding expression measurements that are known to be differentially expressed or control probes that may not match the vsn model, thus avoiding that they influence the normalization process. This operates at the level of probesets, not probes. Both 'subset' and 'subsample' can be used together.
log2scale
If TRUE, this will perform a global affine transform on the data to put them on a similar scale as the original non-transformed data. Many users prefer this. Fold-change estimates are not affected by this transform. In some situations, however, it may be helpful to turn this off, e.g., when comparing independently normalized subsets of the data.
log2asymp
If TRUE, this will perform a global affine transform on the data to make the generalized log (asinh) transform be asymptotically identical to a log base 2 transform. Some people find this helpful. Only one of 'log2scale' or 'log2asymp' can be set to TRUE. Fold-change estimates are not affected by this transform.
...
Further parameters for vsn2.

Value

AffyBatch. The vsn object returned, which can be used as reference for subsequent fits, is provided by description(abatch)@preprocessing$vsnReference.

Details

Please refer to the Details and References sections of the man page for vsn2 for more details about this method.

Important note: after calling vsn2, the function normalize.AffyBatch.vsn exponentiates the data (base 2). This is done in order to make the behavior of this function similar to the other normalization methods in affy. That packages uses the convention of taking the logarithm to base in subsequent analysis steps (e.g. in medpolish).

See Also

vsn2

Examples

Run this code
## Please see vignette.

Run the code above in your browser using DataLab