miRNAss {miRNAss}

MiRNAss: Genome-wide pre-miRNA discovery from few labeled examples

Description

This is the main function of the miRNAss package and implements the miRNA prediction method, It takes as main parameters a matrix with numerical features extracted from RNA hairpins and an incomplent vector of labels where the positive number represents known miRNAs, the negative are not-miRNA hairpins and te zero values are unknown sequences (those that will be classified). As a results it returns a complete label vector.

Usage

miRNAss(sequenceFeatures = NULL, sequenceLabels, AdjMatrix = NULL,
  nNearestNeighbor = 10, missPenalization = 1, scallingMethod = "relief",
  thresholdObjective = "Gm", neg2label = 0.05, positiveProp = NULL,
  eigenVectors = NULL, nEigenVectors = min(400,
  round(length(sequenceLabels)/5)), threadNumber = NA)

Arguments

sequenceFeatures

Data frame with features extracted from stem-loop sequences. It is not required if the adjacency matrix is provided.

sequenceLabels

Vector of labels of the stem-loop sequences. It must have -1 for negative examples, 1 for known miRNAs and zero for the unknown sequences (the ones that would be classificated).

AdjMatrix

Sparse adjacency matrix representeing the graph. If sequence features are provided it is ignored.

nNearestNeighbor

Number of nearest neighbors in the KNN graph. The default value is 10.

missPenalization

Penalization of the missclassification of known examples. The default value is 1. If the examples are not very confident, this value can be diminished.

scallingMethod

Method used for normalization and scalling of the features. The options are 'none', 'whitening' and 'relief' (the default option). The first option does nothing, the second calls the built-in function 'scale' and the last one uses the ReliefFexpRank algorithm from the coreLearn package.

thresholdObjective

Performance measure that would be optimized when estimating the threshold. The options are 'Gm' (geometric mean of the SE and the SP), 'G' (geometric mean of the SE and the precision), 'F1' (harmonic mean between SE and the precision) and 'none' (do not calculate any threshold). The default value is 'Gm'.

neg2label

Proportion of unlabeled stem-loops that would be labeled as negative with the automatic method to start the classification algorithm. The default is 0.05.

positiveProp

Expected proportion of positive sequences. If it is not provided by the user, is estimated as sum(y > 0) / sum(y != 0) when there are negative examples or as 2 * sum(y > 0) / sum(y == 0) when not.

eigenVectors

Eigen decomposition of the Laplacian matrix, as returned by the function eigenDecomposition. If is not provided is calculated internally (this parameter allows to calculate the eigen vectors once and then run several times miRNAss with the same eigen vectors).

nEigenVectors

Number of eigen vectors used to aproximate the solution of the optimization problem. If the number is too low, smoother topographic solutions are founded, probabily losing SP but achieving a better SE. Generally, 400 are enought.

threadNumber

Number of threads used for the calculations. If it is NA leave OpenMP decide the number (may vary across different platforms).

Value

Returns a vector with the same size of the input vector y with the prediction scores for all sequences (even the labelled examples). If a threshold Objective different from 'none' was set, the threshold is estimated and subtracted from the scores, therefore the new threshold that divide the classes is zero. Also, the positive scores are divided by the max positive score, and the negative scores are divided by the magnitud of the minimum negative score.

Examples

# First construct the label vector with the CLASS column
y = as.numeric(celegans$CLASS)*2 - 1

# Remove some labels to make a test
y[sample(which(y>0),200)] = 0
y[sample(which(y<0),700)] = 0

# Take all the features but remove the label column
x = subset(celegans, select = -CLASS)

# Call miRNAss with default parameters
p = miRNAss(x,y)

# Calculate some performance measures
SE = mean(p[ celegans$CLASS & y==0] > 0)
SP = mean(p[!celegans$CLASS & y==0] < 0)
cat("Sensitivity: ", SE, "\nSpecificity: ", SP, "\n")