R/tune.mint.splsda.R

# ========================================================================================================
# tune.mint.splsda: chose the optimal number of parameters per component on a mint.splsda method
# ========================================================================================================
#' Estimate the parameters of mint.splsda method
#' 
#' Computes Leave-One-Group-Out-Cross-Validation (LOGOCV) scores on a
#' user-input grid to determine optimal values for the sparsity parameters in
#' \code{mint.splsda}.
#' 
#' This function performs a Leave-One-Group-Out-Cross-Validation (LOGOCV),
#' where each of \code{study} is left out once. It returns a list of variables
#' of \code{X} that were selected on each of the \code{ncomp} components. Then,
#' a \code{\link{mint.splsda}} can be performed with \code{keepX} set as the
#' output \code{choice.keepX}.
#' 
#' All component \eqn{1:\code{ncomp}} are tuned, except the first ones for
#' which a \code{already.tested.X} is provided. See examples below.
#' 
#' The function outputs the optimal number of components that achieve the best
#' performance based on the overall error rate or BER. The assessment is
#' data-driven and similar to the process detailed in (Rohart et al., 2016),
#' where one-sided t-tests assess whether there is a gain in performance when
#' adding a component to the model. Our experience has shown that in most case,
#' the optimal number of components is the number of categories in \code{Y} -
#' 1, but it is worth tuning a few extra components to check (see our website
#' and case studies for more details).
#' 
#' BER is appropriate in case of an unbalanced number of samples per class as
#' it calculates the average proportion of wrongly classified samples in each
#' class, weighted by the number of samples in each class. BER is less biased
#' towards majority classes during the performance assessment.
#' 
#' More details about the prediction distances in \code{?predict} and the
#' supplemental material of the mixOmics article (Rohart et al. 2017).
#' 
#' @param X numeric matrix of predictors. \code{NA}s are allowed.
#' @param Y Outcome. Numeric vector or matrix of responses (for multi-response
#' models)
#' @param ncomp Number of components to include in the model (see Details).
#' Default to 1
#' @param study grouping factor indicating which samples are from the same
#' study
#' @param test.keepX numeric vector for the different number of variables to
#' test from the \eqn{X} data set
#' @param already.tested.X if \code{ncomp > 1} Numeric vector indicating the
#' number of variables to select from the \eqn{X} data set on the firsts
#' components
#' @param dist only applies to an object inheriting from \code{"plsda"} or
#' \code{"splsda"} to evaluate the classification performance of the model.
#' Should be a subset of \code{"max.dist"}, \code{"centroids.dist"},
#' \code{"mahalanobis.dist"}. Default is \code{"all"}. See
#' \code{\link{predict}}.
#' @param measure Two misclassification measure are available: overall
#' misclassification error \code{overall} or the Balanced Error Rate \code{BER}
#' @param auc if \code{TRUE} calculate the Area Under the Curve (AUC)
#' performance of the model.
#' @param progressBar by default set to \code{TRUE} to output the progress bar
#' of the computation.
#' @param scale Logical. If scale = TRUE, each block is standardized to zero
#' means and unit variances (default: TRUE)
#' @param tol Convergence stopping value.
#' @param max.iter integer, the maximum number of iterations.
#' @param near.zero.var Logical, see the internal \code{\link{nearZeroVar}}
#' function (should be set to TRUE in particular for data with many zero
#' values). Default value is FALSE
#' @param light.output if set to FALSE, the prediction/classification of each
#' sample for each of \code{test.keepX} and each comp is returned.
#' @param signif.threshold numeric between 0 and 1 indicating the significance
#' threshold required for improvement in error rate of the components. Default
#' to 0.01.
#' @return The returned value is a list with components:
#' \item{error.rate}{returns the prediction error for each \code{test.keepX} on
#' each component, averaged across all repeats and subsampling folds. Standard
#' deviation is also output. All error rates are also available as a list.}
#' \item{choice.keepX}{returns the number of variables selected (optimal keepX)
#' on each component.} \item{choice.ncomp}{returns the optimal number of
#' components for the model fitted with \code{$choice.keepX} }
#' \item{error.rate.class}{returns the error rate for each level of \code{Y}
#' and for each component computed with the optimal keepX}
#' 
#' \item{predict}{Prediction values for each sample, each \code{test.keepX} and
#' each comp.} \item{class}{Predicted class for each sample, each
#' \code{test.keepX} and each comp.}
#' @author Florian Rohart, Al J Abadi
#' @seealso \code{\link{mint.splsda}} and http://www.mixOmics.org for more
#' details.
#' @references Rohart F, Eslami A, Matigian, N, Bougeard S, Lê Cao K-A (2017).
#' MINT: A multivariate integrative approach to identify a reproducible
#' biomarker signature across multiple experiments and platforms. BMC
#' Bioinformatics 18:128.
#' 
#' mixOmics article:
#' 
#' Rohart F, Gautier B, Singh A, Lê Cao K-A. mixOmics: an R package for 'omics
#' feature selection and multiple data integration. PLoS Comput Biol 13(11):
#' e1005752
#' @keywords multivariate dplot
#' @export
#' @examples
#' data(stemcells)
#' data = stemcells$gene
#' type.id = stemcells$celltype
#' exp = stemcells$study
#' 
#' res = mint.splsda(X=data,Y=type.id,ncomp=3,keepX=c(10,5,15),study=exp)
#' out = tune.mint.splsda(X=data,Y=type.id,ncomp=2,near.zero.var=FALSE,
#' study=exp,test.keepX=seq(1,10,1))
#' 
#' out$choice.ncomp
#' out$choice.keepX
#' 
#' \dontrun{
#' 
#' out = tune.mint.splsda(X=data,Y=type.id,ncomp=2,near.zero.var=FALSE,
#' study=exp,test.keepX=seq(1,10,1))
#' out$choice.keepX
#' 
#' ## only tune component 2 and keeping 10 genes on comp1
#' out = tune.mint.splsda(X=data,Y=type.id,ncomp=2, study=exp,
#' already.tested.X = c(10),
#' test.keepX=seq(1,10,1))
#' out$choice.keepX
#' 
#' }
tune.mint.splsda <-
    function (X, 
              Y,
              ncomp = 1,
              study,
              test.keepX = c(5, 10, 15),
              already.tested.X,
              dist = c("max.dist", "centroids.dist", "mahalanobis.dist"),
              measure = c("BER", "overall"),
              auc = FALSE,
              progressBar = FALSE,
              scale = TRUE,
              tol = 1e-06,
              max.iter = 100,
              near.zero.var = FALSE,
              light.output = TRUE, # if FALSE, output the prediction and classification of each sample during each folds, on each comp, for each repeat
              signif.threshold=0.01
    )
    {    
        #-- checking general input parameters --------------------------------------#
        #---------------------------------------------------------------------------#
        
        #------------------#
        #-- check entries --#
        if(missing(X))
            stop("'X'is missing", call. = FALSE)
        
        X = as.matrix(X)
        
        if (length(dim(X)) != 2 || !is.numeric(X))
            stop("'X' must be a numeric matrix.", call. = FALSE)
        
        
        # Testing the input Y
        if(missing(Y))
            stop("'Y'is missing", call. = FALSE)
        if (is.null(Y))
            stop("'Y' has to be something else than NULL.", call. = FALSE)
        
        if (is.null(dim(Y)))
        {
            Y = factor(Y)
        }  else {
            stop("'Y' should be a factor or a class vector.", call. = FALSE)
        }
        
        if (nlevels(Y) == 1)
            stop("'Y' should be a factor with more than one level", call. = FALSE)
        
        #-- check significance threshold
        signif.threshold <- .check_alpha(signif.threshold)
        
        #-- progressBar
        if (!is.logical(progressBar))
            stop("'progressBar' must be a logical constant (TRUE or FALSE).", call. = FALSE)
        
        
        if (is.null(ncomp) || !is.numeric(ncomp) || ncomp <= 0)
            stop("invalid number of variates, 'ncomp'.")
        
        
        #-- measure
        measure <- match.arg(measure)
        
        #if ((!is.null(already.tested.X)) && (length(already.tested.X) != (ncomp - 1)) )
        #stop("The number of already tested parameters should be NULL or ", ncomp - 1, " since you set ncomp = ", ncomp)
        if (missing(already.tested.X))
        {
            already.tested.X = NULL
        } else {
            if(is.list(already.tested.X))
                stop("''already.tested.X' must be a vector of keepX values")
            
            message(paste("Number of variables selected on the first", length(already.tested.X), "component(s):", paste(already.tested.X,collapse = " ")))
        }
        if(length(already.tested.X) >= ncomp)
            stop("'ncomp' needs to be higher than the number of components already tuned, which is length(already.tested.X)=",length(already.tested.X) , call. = FALSE)
        
        
        # -- check using the check of mint.splsda
        Y.mat = unmap(Y)
        colnames(Y.mat) = levels(Y)
        
        check = Check.entry.pls(X, Y = Y.mat, ncomp = ncomp, mode="regression", scale=scale,
                                near.zero.var=near.zero.var, max.iter=max.iter ,tol=tol ,logratio="none" ,DA=TRUE, multilevel=NULL)
        X = check$X
        ncomp = check$ncomp
        
        
        # -- study
        #set the default study factor
        if (missing(study))
            stop("'study' is missing", call. = FALSE)
        
        if (length(study) != nrow(X))
            stop(paste0("'study' must be a factor of length ",nrow(X),"."))
        
        if (any(table(study) <= 1))
            stop("At least one study has only one sample, please consider removing before calling the function again", call. = FALSE)
        if (any(table(study) < 5))
            warning("At least one study has less than 5 samples, mean centering might not do as expected")
        
        if(sum(apply(table(Y,study)!=0,2,sum)==1) >0)
            stop("At least one study only contains a single level of the multi-levels outcome Y. The MINT algorithm cannot be computed.")
        
        if(sum(apply(table(Y,study)==0,2,sum)>0) >0)
            warning("At least one study does not contain all the levels of the outcome Y. The MINT algorithm might not perform as expected.")
        
        
        #-- dist
        dist = match.arg(dist)
        
        #-- light.output
        if (!is.logical(light.output))
            stop("'light.output' must be either TRUE or FALSE", call. = FALSE)
        
        #-- test.keepX
        if (is.null(test.keepX) | length(test.keepX) == 1 | !is.numeric(test.keepX))
            stop("'test.keepX' must be a numeric vector with more than two entries", call. = FALSE)
        
        #-- end checking --#
        #------------------#
        
        
        
        #-- cross-validation approach  ---------------------------------------------#
        #---------------------------------------------------------------------------#
        test.keepX = sort(test.keepX) #sort test.keepX so as to be sure to chose the smallest in case of several minimum
        
        # if some components have already been tuned (eg comp1 and comp2), we're only tuning the following ones (comp3 comp4 .. ncomp)
        if ((!is.null(already.tested.X)))
        {
            comp.real = (length(already.tested.X) + 1):ncomp
        } else {
            comp.real = 1:ncomp
        }
        
        mat.error = matrix(nrow = length(test.keepX), ncol = 1,
                           dimnames = list(test.keepX,1))
        rownames(mat.error) = test.keepX
        
        error.per.class = list()
        
        mat.sd.error = matrix(0,nrow = length(test.keepX), ncol = ncomp-length(already.tested.X),
                              dimnames = list(c(test.keepX), c(paste0('comp', comp.real))))
        mat.mean.error = matrix(nrow = length(test.keepX), ncol = ncomp-length(already.tested.X),
                                dimnames = list(c(test.keepX), c(paste0('comp', comp.real))))
        
        error.per.class.mean = matrix(nrow = nlevels(Y), ncol = ncomp-length(already.tested.X),
                                      dimnames = list(c(levels(Y)), c(paste0('comp', comp.real))))
        error.per.class.sd = matrix(0,nrow = nlevels(Y), ncol = ncomp-length(already.tested.X),
                                    dimnames = list(c(levels(Y)), c(paste0('comp', comp.real))))
        
        
        error.per.study.keepX.opt = matrix(nrow = nlevels(study), ncol = ncomp-length(already.tested.X),
                                           dimnames = list(c(levels(study)), c(paste0('comp', comp.real))))
        
        if(light.output == FALSE)
            prediction.all = class.all = list()
        if(auc)
            auc.mean=list()
        
        error.per.class.keepX.opt=list()
        
        # successively tune the components until ncomp: comp1, then comp2, ...
        for(comp in 1:length(comp.real))
        {
            
            if (progressBar == TRUE)
                cat("\ncomp",comp.real[comp], "\n")
            
            result = LOGOCV (X, Y, ncomp = 1 + length(already.tested.X), study = study,
                             choice.keepX = already.tested.X,
                             test.keepX = test.keepX, measure = measure,
                             dist = dist, near.zero.var = near.zero.var, progressBar = progressBar, scale = scale, max.iter = max.iter, auc = auc)
            
            
            # in the following, there is [[1]] because 'tune' is working with only 1 distance and 'MCVfold.spls' can work with multiple distances
            mat.mean.error[, comp]=result[[measure]]$error.rate.mean[[1]]
            if (!is.null(result[[measure]]$error.rate.sd[[1]]))
                mat.sd.error[, comp]=result[[measure]]$error.rate.sd[[1]]
            
            # confusion matrix for keepX.opt
            error.per.class.keepX.opt[[comp]]=result[[measure]]$confusion[[1]]
            
            # best keepX
            already.tested.X = c(already.tested.X, result[[measure]]$keepX.opt[[1]])
            
            # error per study for keepX.opt
            error.per.study.keepX.opt[,comp] = result[[measure]]$error.per.study.keepX.opt[[1]]
            
            if(light.output == FALSE)
            {
                #prediction of each samples for each fold and each repeat, on each comp
                class.all[[comp]] = result$class.comp[[1]]
                prediction.all[[comp]] = result$prediction.comp
            }
            if(auc)
                auc.mean[[comp]]=result$auc
            
        } # end comp
        names(error.per.class.keepX.opt) = c(paste0('comp', comp.real))
        names(already.tested.X) = c(paste0('comp', 1:ncomp))
        
        if (progressBar == TRUE)
            cat('\n')
        
        # calculating the number of optimal component based on t.tests and the error.rate.all, if more than 3 error.rates(repeat>3)
        if(nlevels(study) > 2 & length(comp.real) >1)
        {
            opt = t.test.process(error.per.study.keepX.opt, alpha=signif.threshold)
            ncomp_opt = comp.real[opt]
        } else {
            ncomp_opt = NULL
        }
        
        result = list(
            error.rate = mat.mean.error,
            choice.keepX = already.tested.X,
            choice.ncomp = list(ncomp = ncomp_opt, values = error.per.study.keepX.opt),
            error.rate.class = error.per.class.keepX.opt)
        
        if(auc)
        {
            names(auc.mean) = c(paste0('comp', comp.real))
            result$auc = auc.mean
        }
        
        if(light.output == FALSE)
        {
            names(class.all) = names(prediction.all) = c(paste0('comp', comp.real))
            result$predict = prediction.all
            result$class = class.all
        }
        result$measure = measure
        result$call = match.call()
        
        class(result) = c("tune.mint.splsda","tune.splsda")
        
        return(result)
    }
mixOmicsTeam/mixOmics documentation built on Nov. 4, 2024, 8:56 a.m.