pamr.train {pamr}

A function to train a nearest shrunken centroid classifier

Description

A function that computes a nearest shrunken centroid for gene expression (microarray) data

Usage

pamr.train(
  data,
  gene.subset = NULL,
  sample.subset = NULL,
  threshold = NULL,
  n.threshold = 30,
  scale.sd = TRUE,
  threshold.scale = NULL,
  se.scale = NULL,
  offset.percent = 50,
  hetero = NULL,
  prior = NULL,
  remove.zeros = TRUE,
  sign.contrast = "both",
  ngroup.survival = 2
)

Arguments

data

The input data. A list with components: x- an expression genes in the rows, samples in the columns), and y- a vector of the class labels for each sample. Optional components- genenames, a vector of gene names, and geneid- a vector of gene identifiers.

gene.subset

Subset of genes to be used. Can be either a logical vector of length total number of genes, or a list of integers of the row numbers of the genes to be used

sample.subset

Subset of samples to be used. Can be either a logical vector of length total number of samples, or a list of integers of the column numbers of the samples to be used.

threshold

A vector of threshold values for the centroid shrinkage.Default is a set of 30 values chosen by the software

n.threshold

Number of threshold values desired (default 30)

scale.sd

Scale each threshold by the wthin class standard deviations? Default: true

threshold.scale

Additional scaling factors to be applied to the thresholds. Vector of length equal to the number of classes. Default- a vectors of ones.

se.scale

Vector of scaling factors for the within class standard errors. Default is sqrt(1/n.class-1/n), where n is the overall sample size and n.class is the sample sizes in each class. This default adjusts for different class sizes.

offset.percent

Fudge factor added to the denominator of each t-statistic, expressed as a percentile of the gene standard deviation values. This is a small positive quantity to penalize genes with expression values near zero, which can result in very large ratios. This factor is expecially impotant for Affy data. Default is the median of the standard deviations of each gene.

hetero

Should a heterogeneity transformation be done? If yes, hetero must be set to one of the class labels (see Details below). Default is no (hetero=NULL)

prior

Vector of length the number of classes, representing prior probabilities for each of the classes. The prior is used in Bayes rule for making class prediction. Default is NULL, and prior is then taken to be n.class/n, where n is the overall sample size and n.class is the sample sizes in each class.

remove.zeros

Remove threshold values yielding zero genes? Default TRUE

sign.contrast

Directions of allowed deviations of class-wise average gene expression from the overall average gene expression. Default is ⁠both'' (positive or negative). Can also be set to ⁠positive” or “negative”.

ngroup.survival

Number of groups formed for survival data. Default 2

Details

pamr.train fits a nearest shrunken centroid classifier to gene expression data. Details may be found in the PNAS paper referenced below. One feature not described there is "heterogeneity analysis". Suppose there are two classes labelled "A" and "B". CLass "A" is considered a normal class, and "B" an abnormal class. Setting hetero="A" transforms expression values x[i,j] to |x[i,j]- mean(x[i,j])| where the mean is taken only over samples in class "A". The transformed feature values are then used in Pam. This is useful when the abnormal class "B" is heterogeneous, i.e. a given gene might have higher expresion than normal for some class "B" samples, and lower for others. With more than 2 classes, each class is centered on the class specified by hetero.

Value

A list with components

y

The outcome classes.

yhat

A matrix of predicted classes, each column representing the results from one threshold.

.

prob

A array of predicted class probabilities. of dimension n by nclass by n.threshold. n is the number samples, nclass is the number of classes, n.threshold is the number of thresholds tried

centroids

A matrix of (unshrunken) class centroids, n by nclass

hetero

Value of hetero used in call to pamr.train

norm.cent

Centroid of "normal" group, if hetero was specified

centroid.overall

A vector containing the (unshrunken) overall centroid (all classes together)

sd

A vector of the standard deviations for each gene

threshold

A vector of the threshold tried in the shrinkage

nonzero

A vector of the number of genes that survived the thresholding, for each threshold value tried

threshold.scale

A vector of threshold scale factors that were used

se.scale

A vector of standard error scale factors that were used

call

The calling sequence used

prior

The prior probabilities used

errors

The number of trainin errors for each threshold value

Author(s)

Trevor Hastie,Robert Tibshirani, Balasubramanian Narasimhan, and Gilbert Chu

References

Robert Tibshirani, Trevor Hastie, Balasubramanian Narasimhan, and Gilbert Chu Diagnosis of multiple cancer types by shrunken centroids of gene expression PNAS 99: 6567-6572. Available at www.pnas.org

Examples

#generate some data
suppressWarnings(RNGversion("3.5.0"))
set.seed(120)
x <- matrix(rnorm(1000*20),ncol=20)
y <- sample(c(1:4),size=20,replace=TRUE)
mydata <- list(x=x,y=factor(y))

#train classifier
results<-   pamr.train(mydata)

# train classifier on all  data except class 4
results2 <- pamr.train(mydata,sample.subset=(mydata$y!=4))
 
# train classifier on  only the first 500 genes
results3 <- pamr.train(mydata,gene.subset=1:500)