R/model_specific_feature_selection.R

In ahmedelmahy/RNASeqclassifier1: A collection of functions for feature selection and classifier development from RNASeq data.

#' Eliminate features with Random forest by sorting variable importance.
#'
#' @param x train dataframe
#' @param y train class
#' @param keepn Number of genes to keep
#' @param fitControl a configuration list for caret package
#' @return a vector of selected genes
eliminate_RF <- function(x,y, keepn = 100, fitControl){
  fit_tune <- train(x = x,
                    y = y,
                    method = "rf",
                    trControl = fitControl)

  fit <- caret::train(x = x,
                      y = y,
                      method ="rf",
                      tuneGrid=expand.grid(.mtry=fit_tune$bestTune$mtry),
                      trControl = fitControl)
  var_imp <- varImp(fit) # random forest
  var_imp <- data.frame(cbind(row.names(var_imp$importance),
                              var_imp$importance[, 1]))
  names(var_imp) <- c("genes", "imp")
  var_imp <- var_imp[order(as.numeric(as.matrix(var_imp$imp)),
                           decreasing = T), ]
  topgenes <- as.character(var_imp$genes[1:keepn])
  return(topgenes)
}