switchBox (version 1.6.0)

SWAP.KTSP.Train: Function for training the K-TSP classifier.

Description

SWAP.KTSP.Train trains a binary K-TSP classifier. The classifiers resulting from using this function can be passed to SWAP.KTSP.Classify for samples classification.

Usage

SWAP.KTSP.Train(inputMat, phenoGroup, krange = c(3, 5, 7:10), FilterFunc = SWAP.Filter.Wilcoxon, RestrictedPairs, ...)

Arguments

inputMat

is a numerical matrix containing the measurements (e.g., gene expression data) to be used to build the K-TSP classifier. The columns represent samples and the rows represent the features (e.g., genes). The number of columns must agree with the length of phenoGroup. Note that rownames(inputMat) will be used as the feature names (e.g., gene symbols) in all subsequent analyses.

phenoGroup

is a factor with two levels containing the phenotype information used to train the K-TSP classifier. In order to identify the best TSP to be included in the classifier, the features contained in inputMat will be compared between the two groups defined by this factor. Levels from phenoGroup will be also used to reorder the features in each TSP such as the first feature is larger than the second one in the group corresponding to first level, and vice-versa.

krange

an integer (or a vector of integers) defining the candidate number of Top Scoring Pairs (TSPs) from which the algorithm chooses to build the final classifier. The algorithm uses the mechanism in Afsari et al (AOAS, 2014) to select the number of pairs and pair of features.

FilterFunc

is a filtering function to reduce the starting number of features to be used to identify the Top Scoring Pairs (TSP). The default filter is differential expression test based on the Wilcoxon rank-sum test and alternative filtering functions can be passed too (see SWAP.Filter.Wilcoxon for details). The output of the function must be subset of rownames(inputMat)

RestrictedPairs

is a character matrix with two columns containing the feature pairs to be considered for score calculations. Each row should contain a pair of feature names matching the rownames of inputMat. If RestrictedPairs is missing all available feature pairs will be considered.

...

Additional argument passed to the filtering function FilterFunc.

Value

The KTSP classifier, in the form of a list, which contains the following components:

name

The classifier name.

TSPs

A k by 2 matrix, containing the feature names for each TSP. These names correspond to the rownames(inputData). In this matrix each row corresponds to a specific TSP. For each TSP (i.e. row in the TSPs matrix) the order of the features is such that the first one is on average smaller than the second one in the phenotypic group defined by the first levels of the phenoGroup factor and vice-versa. The algorithm uses the mechanism in Afsari et al (2014) to select the number of pairs and pair of features.

$score

scores TSP for the top k TSPs.

$label

The class labels. These labels correspond to the phenoGroup factor lelves and will be used lable any new sample classified by the SWAP.KTSP.Classify function.

References

See switchBox for the references.

See Also

SWAP.KTSP.Classify, SWAP.Filter.Wilcoxon, SWAP.CalculateSignedScore

Examples

##################################################
### Load gene expression data for the training set
data(trainingData)


### Show group variable for the TRAINING set
table(trainingGroup)


##################################################
### Train a classifier using default filtering function based on the Wilcoxon test
classifier <- SWAP.KTSP.Train(matTraining, trainingGroup, krange=c(3, 5, 8:15))

### Show the classifier
classifier


##################################################
### Train another classifier from the top 4 best features 
### according to the deafault  filtering function
classifier <- SWAP.KTSP.Train(matTraining, trainingGroup,
			      FilterFunc=SWAP.Filter.Wilcoxon, featureNo=4)

### Show the classifier
classifier


##################################################
### To use all features "FilterFunc" must be set to NULL
classifier <- SWAP.KTSP.Train(matTraining, trainingGroup, FilterFunc=NULL)

### Show the classifier
classifier


##################################################
### Train a classifier using and alternative filtering function.
### For instance we can use the a  "t.test"  to selec the features 
### with an absolute t-statistics larger than a specified quantile
topRttest <- function(situation, data, quant = 0.75) {
	out <- apply(data, 1, function(x, ...) t.test(x ~ situation)$statistic )
	names(out[ abs(out) > quantile(abs(out), quant) ])
}

### Show the top features selected 
topRttest(trainingGroup, matTraining, quant=0.95)

### Train a classifier using the alternative filtering function
### and also define the maximum number of TSP using "krange"
classifier <- SWAP.KTSP.Train(matTraining, trainingGroup,
			      FilterFunc = topRttest, quant = 0.75, krange=c(15:30) )

### Show the classifier
classifier


##################################################
### Training with restricted pairs

### Define a set of specific pairs to be used for classifier development
### For this example we will a random set of features
### In a real example these pairs should  be provided by the user.
set.seed(123)
somePairs <- matrix(sample(rownames(matTraining), 6^2, replace=FALSE), ncol=2)
head(somePairs, n=3)
dim(somePairs)

### Train a classifier using the restricted feature pairs and the default filtering
classifier <- SWAP.KTSP.Train(matTraining, trainingGroup,
			      RestrictedPairs = somePairs, krange=3:16)

### Show the classifier
classifier