R/sharedFunctions.R

In trena: Fit transcriptional regulatory networks using gene expression, priors, machine learning

#----------------------------------------------------------------------------------------------------
#' Run the Elastic Net Solvers
#' @description Given a TReNA object with either LASSO or Ridge Regression as the solver, use the \code{\link{glmnet}} function to estimate coefficients for each transcription factor as a predictor of the target gene's expression level.
#'
#' @param obj An object of class Solver
#' @param target.gene A designated target gene that should be part of the mtx.assay data
#' @param tfs The designated set of transcription factors that could be associated with the target gene.
#' @param tf.weights A set of weights on the transcription factors (default = rep(1, length(tfs)))
#' @param alpha The LASSO/Ridge tuning parameter
#' @param lambda The penalty tuning parameter for elastic net
#' @param keep.metrics A binary variable indicating whether or not to keep metrics
#'
#' @return A data frame containing the coefficients relating the target gene to each transcription factor, plus other fit parameters
#'
#' @seealso \code{\link{glmnet}}
#'
#'

elasticNetSolver <-  function (obj, target.gene, tfs, tf.weights, alpha, lambda, keep.metrics){

    if(length(tfs) == 0)
        return(data.frame())

    # we don't try to handle tf self-regulation
    deleters <- grep(target.gene, tfs)
    if(length(deleters) > 0){
        tfs <- tfs[-deleters]
        tf.weights <- tf.weights[-deleters]
        if(!obj@quiet)
            message(sprintf("Removing target.gene from candidate regulators: %s", target.gene))
    }

    if( length(tfs) == 0 ) return( data.frame() )

    mtx <- getAssayData(obj)
    stopifnot(target.gene %in% rownames(mtx))
    stopifnot(all(tfs %in% rownames(mtx)))
    stopifnot(class(lambda) %in% c("NULL","numeric"))
    features <- t(mtx[tfs,,drop=FALSE ])
    target <- as.numeric(mtx[target.gene,])

    if( length(tfs) == 1 ) {
        fit = stats::lm( target ~ features )
        mtx.beta = stats::coef(fit)
        mtx.beta = data.frame( beta = mtx.beta[2] , intercept = mtx.beta[1])
        rownames(mtx.beta) = tfs
        return(mtx.beta)
          # branching on keep.metrix disabled: cor.target.feature is neither defined nor assigned
          # if( keep.metrics == FALSE ) return( mtx.beta )
          #  if( keep.metrics == TRUE ) return( list( mtx.beta = mtx.beta , lambda = NA , r2 = cor.target.feature^2 ) )
    }

    if( length(lambda) == 0  ) {

        # Run Permutation testing to find lambda
        if( alpha != 0 )
            alpha.perm = alpha
        else(alpha.perm = 0.1)
        target.mixed <- sample(target)
        threshold <- 1E-15
        lambda.change <- 10^(-4)
        lambda <- 1
        lambda.list <- numeric(length=50)

        for(i in 1:length(lambda.list)){
            # Do a binary search
            step.size <- lambda/2 # Start at 0.5
            while(step.size > lambda.change){

                # Get the fit
                fit <- glmnet(features, target.mixed, penalty.factor = tf.weights, alpha=alpha.perm, lambda=lambda)

                # Case 1: nonsense, need to lower lambda
                if(max(fit$beta) < threshold){
                    lambda <- lambda - step.size
                }

                # Case 2: sense, need to raise lambda
                else{
                    lambda <- lambda + step.size
                }
                # Halve the step size and re-scramble the target
                step.size <- step.size/2
                target.mixed <- sample(target)
            }
            lambda.list[[i]] <- lambda
        }
        # Give lambda as 1 + 1se
        lambda <- mean(lambda.list) + (stats::sd(lambda.list)/sqrt(length(lambda.list)))

        fit <- glmnet(features, target, penalty.factor=tf.weights, alpha=alpha, lambda=lambda)

    }

    # For non-LASSO
    #        else{
    #            fit <- cv.glmnet(features, target, penalty.factor=tf.weights, grouped=FALSE , alpha = alpha )
    #            lambda.min <- fit$lambda.min
    #            lambda <-fit$lambda.1se
    #        }

    else if(is.numeric(lambda)){
        fit = glmnet(features, target, penalty.factor=tf.weights, alpha=alpha, lambda=lambda)
    }

    # extract the exponents of the fit
    mtx.beta <- as.matrix( stats::predict( fit , newx = features , type = "coef" , s = lambda ) )
    colnames(mtx.beta) <- "beta"
    deleters <- as.integer(which(mtx.beta[,1] == 0))
    if( all( mtx.beta[,1] == 0 ) ) return( data.frame() )
    if(length(deleters) > 0)
        mtx.beta <- mtx.beta[-deleters, , drop=FALSE]

    # put the intercept, admittedly with much redundancy, into its own column
    intercept <- mtx.beta[1,1]
    mtx.beta <- mtx.beta[-1, , drop=FALSE]
    mtx.beta <- cbind(mtx.beta, intercept=rep(intercept, nrow(mtx.beta)))

    #if(!obj@quiet)
    #   graphics::plot(fit.nolambda, xvar='lambda', label=TRUE)

    if( nrow(mtx.beta) > 1 ) {
        ordered.indices <- order(abs(mtx.beta[, "beta"]), decreasing=TRUE)
        mtx.beta <- mtx.beta[ordered.indices,]
    }

    mtx.beta <- as.data.frame(mtx.beta)

    if( keep.metrics == TRUE ) {
        pred.values <- stats::predict( fit , newx = features , s = lambda , type = "link" )
        r2 <- (stats::cor( target , pred.values )[1,1])^2
        return( list( mtx.beta = mtx.beta , lambda = lambda , r2 = r2 ) )
    }

    if( keep.metrics == FALSE )
        return(mtx.beta)
} # elasticNetSolver
#----------------------------------------------------------------------------------------------------