R/plottingBayes.R

In pRoloc: A unifying bioinformatics framework for spatial proteomics

##' Produces a pca plot with uncertainty in organelle means projected
##' onto the PCA plot with contours.
##'
##' @title Uncertainty plot organelle means
##' @param object A valid object of class \code{MSnset}
##' @param params A valid object of class \code{MCMCParams} that has been
##'     processed and checked for convergence
##' @param priors The prior that were used in the model
##' @param dims The PCA dimension in which to project he data, default is
##'     \code{c(1,2)}
##' @param fcol The columns of the feature data which contain the marker data.
##' @param aspect A argument to change the plotting aspect of the PCA
##' @return Used for side effect of producing plot. Invisibily returns an ggplot
##'     object that can be further manipulated
##' @author Oliver M. Crook <omc25@cam.ac.uk>
##' @examples
##' \dontrun{
##' library("pRolocdata")
##' data("tan2009r1")
##'
##' tanres <- tagmMcmcTrain(object = tan2009r1)
##' tanres <- tagmMcmcProcess(tanres)
##' tan2009r1 <- tagmMcmcPredict(object = tan2009r1, params = tanres, probJoint = TRUE)
##' myparams <- chains(e14Tagm_converged_pooled)[[1]]
##' myparams2 <- chains(mcmc_pool_chains(tanres))[[1]]
##' priors <- tanres@priors
##' pRoloc:::nicheMeans2D(object = tan2009r1, params = myparams2, priors = priors)
##' }
nicheMeans2D <- function(object,
                         params,
                         priors,
                         dims = c(1, 2),
                         fcol = "markers",
                         aspect = 0.5) {
    ## Undefined global variables
    X1 <- X2 <- organelle <- NULL
    ## Make coordinates
    .pca <- plot2D(object, dims = dims, plot = FALSE)
    d <- 2
    mcmcIter <- seq.int(by = 10, from = 1, to = params@n)
    iter_len <- length(mcmcIter)
    orgMeans <- array(NA, c(params@K, iter_len, ncol(object)))

    ## Compute means at each iteration of MCMC algorithm
    idx <- rownames(params@Component)
    mydata_sub <- object[idx, ]
    for (i in seq.int(mcmcIter)) {
        for (j in seq.int(params@K)) {
            idxj <- (params@Component[, mcmcIter[i]] ==  j)

            mj <- colMeans(rbind(exprs(mydata_sub)[idxj, ],
                                 exprs(object[fData(object)[, fcol] == getMarkerClasses(object)[j], ])) )
            nj <- sum(fData(object)[, "tagm.mcmc.allocation"] == getMarkerClasses(object)[j])
            lambdaj <- params@ComponentParam@lambdak[j]
            orgMeans[j, i, ] <- (nj * mj + priors$lambda0 * priors$mu0)/lambdaj
        }
    }

    ## Get data to align to
    M <- matrix(NA, nrow = params@K, ncol = ncol(object))
    rownames(M) <- getMarkerClasses(object)
    for (j in seq.int(params@K)) {
        M[j, ] <- colMeans(exprs(object)[fData(object)[, fcol] == getMarkerClasses(object)[j],])
    }

    ## Create coordinates
    coords <- matrix(NA, nrow = params@K, ncol = 2)
    res0 <- prcomp(M, scale. = TRUE)
    eigs <- colnames(.pca)

    ## Create inital dataset, computing average location in PCA plot
    Y0 <- matrix(NA, nrow = params@K, ncol = ncol(object))
    for (j in seq.int(params@K)) {
        Y0[j, ] <- colMeans(.pca[fData(object)[, fcol] == getMarkerClasses(object)[j], seq_len(d)])
    }
    Y0.df <- data.frame(organelle = getMarkerClasses(object), Y0)

    ## Repeat for different monte-carlo samples
    Y.lst <- list()
    for ( i in seq.int(iter_len)) {
        res <- prcomp(orgMeans[, i, ], scale = TRUE, center = TRUE)
        res.proc <- vegan::procrustes(Y0, res$x[, dims])
        Y.df <- data.frame(organelle = getMarkerClasses(object), res.proc$Yrot)
        Y.lst[[i]] <- Y.df
    }

    ## create long data formats
    names(Y.lst) <- seq_len(iter_len)
    Y.lst.df <- plyr::ldply(Y.lst, .fun = function(x) x, .id = "mcmcIter")
    table(Y.lst.df$organelle)
    cols <- getStockcol()[1:params@K] # appropriate colours

    ## ggplot
    gg <- ggplot(
        data = dplyr::mutate(Y.lst.df, organelle = factor(organelle)),
        aes(x = X1, y = X2, color = organelle)) +
        coord_fixed() +
        geom_density2d(contour = TRUE) +
        geom_point(alpha = 0.3) +
        xlab(paste0(eigs[1])) +
        ylab(paste0(eigs[2])) +
        theme(legend.position = "right",
              text = element_text(size = 12)) +
        scale_color_manual(values = cols) +
        scale_fill_manual(values = cols) +
        theme_minimal() +
        theme(panel.grid.major = element_blank(),
              panel.grid.minor = element_blank(),
              aspect.ratio = aspect,
              panel.border = element_rect(colour = "black", fill = NA, size = 1),
              plot.title = element_text(hjust = 0.5, size = 20),
              legend.text = element_text(size = 14)) +
        ggtitle(label = "Uncertainty in mean of organelle localisation")
    gg

    return(gg)
}
##' Produces a pca plot with spatial variation in localisation probabilities
##'
##' @title Uncertainty plot in localisation probabilities
##' @param object A valid object of class \code{MSnset} with mcmc prediction
##'     results from \code{tagmMCMCpredict}
##' @param dims The PCA dimension in which to project he data, default is
##'     \code{c(1,2)}
##' @param cov.function The covariance function used default is
##'     wendland.cov. See \code{fields} package.
##' @param theta A hyperparameter to the covariance function. See \code{fields}
##'     package. Default is 1.
##' @param derivative The number of derivative of the wendland kernel. See
##'     \code{fields} package. Default is 2.
##' @param k A hyperparamter to the covariance function. See \code{fields}
##'     package. Default is 1.
##' @param breaks Probability values at which to draw the contour bands. Default
##'     is \code{c(0.99, 0.95, 0.9, 0.85, 0.8, 0.75, 0.7)}
##' @param aspect A argument to change the plotting aspect of the PCA
##' @return Used for side effect of producing plot. Invisibily returns an ggplot
##'     object that can be further manipulated
##' @author Oliver M. Crook <omc25@cam.ac.uk>
##' @examples
##' \dontrun{
##' library("pRolocdata")
##' data("tan2009r1")
##'
##' tanres <- tagmMcmcTrain(object = tan2009r1)
##' tanres <- tagmMcmcProcess(tanres)
##' tan2009r1 <- tagmMcmcPredict(object = tan2009r1, params = tanres, probJoint = TRUE)
##' spatial2D(object = tan2009r1)
##' }
spatial2D <- function(object,
                      dims = c(1, 2),
                      cov.function = fields::wendland.cov,
                      theta = 1,
                      derivative = 2,
                      k = 1,
                      breaks = c(0.99, 0.95, 0.9, 0.85, 0.8, 0.75, 0.7),
                      aspect = 0.5) {
    ## Undefined global variables

    x <- y <- z <- level <- organelle <- NULL
    ## This function requieres three suggested packages. Testing if they are
    ## available.
    missing_packages <- character()
    if (!requireNamespace("vegan"))
        missing_packages <- c(missing_packages, "vegan")
    if (!requireNamespace("fields"))
        missing_packages <- c(missing_packages, "fields")
    if (!requireNamespace("akima"))
        missing_packages <- c(missing_packages, "akima")
    if (length(missing_packages))
        stop("Please install the following package(s) to use this function:\n",
             paste(missing_packages, collapse = ", "))

    ## generate pca plot and create data from with probabilties
    .pca <- plot2D(object, dims = dims, plot = FALSE)
    probs <- data.frame(x = .pca[, 1], y = .pca[, 2], mcmc.prob = fData(object)$tagm.mcmc.joint)
    colnames(probs) <- c(c("x", "y"), getMarkerClasses(object))
    eigs <- colnames(.pca)

    ## put data in appropriate long format
    probs.lst <- list()
    for(j in getMarkerClasses(object)) {
        probs.lst[[j]] <- probs[, c("x", "y", j)]
        colnames(probs.lst[[j]]) <- c("x", "y", "probability")
    }
    probs.lst.df <- plyr::ldply(probs.lst, .fun = function(x) x, .id = "organelle")

    ## Create storage
    coords <- list()
    locations <- list()
    df <- list()
    ## Create appropriate spatial grid
    for (j in getMarkerClasses(object)) {
        idxOrg <- c(probs.lst.df$organelle == j)
        coords[[j]] <- akima::interp(x = probs.lst.df$x[idxOrg],
                                     y = probs.lst.df$y[idxOrg],
                                     z = probs.lst.df$probability[idxOrg],
                                     extrap=FALSE, linear = TRUE, duplicate = TRUE) # interpolate onto appropriate grid
        coords[[j]]$z[is.na(coords[[j]]$z)] <- 0 # NaNs beyond data set to 0
        locations[[j]] <- cbind(rep(coords[[j]]$x, 40), rep(coords[[j]]$y, each = 40)) # get grid
        smoothedprobs <- fields::smooth.2d(coords[[j]]$z, x = locations[[j]],
                                           cov.function = cov.function,
                                           theta = theta,
                                           derivative = derivative, k = k) # spatial smoothing of probabilities
        ## normalisation and formatting
        zmat <- matrix(smoothedprobs$z, ncol = 1)
        zmat <- zmat/max(zmat)
        df[[j]] <- data.frame(x = rep(smoothedprobs$x, 64), y = rep(smoothedprobs$y, each = 64), z = zmat)
    }
    ## format data
    df.lst <- plyr::ldply(df, .fun = function(x) x, .id = "organelle")
    df.lst <- dplyr::mutate(df.lst, organelle = factor(organelle))
    K <- length(getMarkerClasses(object))
    cols <- getStockcol()[1:K] # get appropriate colours


    gg <- ggplot(
        data = df.lst,
        aes(x = x, y = y, z = z, color = organelle)) +
        coord_fixed() +
        geom_contour(breaks = breaks, size = 1.2, aes(alpha = ggplot2::stat(level))) +
        geom_point(alpha = 0) +
        xlab(paste0(eigs[1])) +
        ylab(paste0(eigs[2])) +
        scale_alpha(guide = "none") +
        theme(legend.position = "right",
              text = element_text(size = 12)) +
        scale_color_manual(values = cols) +
        scale_fill_manual(values = cols) +
        theme_minimal() +
        theme(panel.grid.major = element_blank(), panel.grid.minor = element_blank(),
              aspect.ratio = aspect,
              panel.border = element_rect(colour = "black", fill = NA, size = 1),
              plot.title = element_text(hjust = 0.5, size = 20),
              legend.text=element_text(size = 14)) +
        ggtitle(label = "Spatial variation of localisation probabilities")

    return(gg)
}



##' Model calibration model with posterior z-scores and posterior shrinkage
##'
##' @title Model calibration plots
##' @param object A valid object of class \code{MSnset}
##' @param params A valid object of class \code{MCMCParams} that has been processed
##' and checked for convergence
##' @param priors The prior that were used in the model
##' @param fcol The columns of the feature data which contain the marker data.
##' @return Used for side effect of producing plot. Invisibily returns an ggplot object
##' that can be further manipulated
##' @author Oliver M. Crook <omc25@cam.ac.uk>
##' @examples
##' \dontrun{
##' library("pRoloc")
##' data("tan2009r1")
##'
##' tanres <- tagmMcmcTrain(object = tan2009r1)
##' tanres <- tagmMcmcProcess(tanres)
##' tan2009r1 <- tagmMcmcPredict(object = tan2009r1, params = tanres, probJoint = TRUE)
##' myparams <- chains(e14Tagm_converged_pooled)[[1]]
##' myparams2 <- chains(mcmc_pool_chains(tanres))[[1]]
##' priors <- tanres@priors
##' pRoloc:::mixing_posterior_check(object = tan2009r1, params = myparams2, priors = priors)
##' }
mixing_posterior_check <- function(object,
                                   params,
                                   priors,
                                   fcol = "markers") {
    ## Undefined global variables
    posteriorShrinkage <- posteriorZscore <- organelle <- NULL

    K <- length(priors$beta0)
    N <- nrow(object)

    ## Compute prior mean and variance
    tallyComp <- c(table(fData(object)[, fcol])[1:K] + priors$beta0)
    meanComp <- tallyComp/sum(tallyComp)
    varComp <- meanComp * (1 - meanComp)/sum(tallyComp + 1)

    ## Compute posterior quantities
    cmptable <- apply(params@Component, 2, tabulate)
    varCmptable <- apply(cmptable/N, 1, var)
    sdCmptable <- apply(cmptable/N, 1, sd)

    post_z_mixing <- abs((rowMeans(cmptable/N) - meanComp)/sdCmptable)
    post_shrink_mixing <- 1 - (varCmptable/varComp)^2

    post_check_mixing <- data.frame(x = post_shrink_mixing, y = post_z_mixing, getMarkerClasses(object))
    colnames(post_check_mixing) <- c("posteriorShrinkage", "posteriorZscore", "organelle")
    rownames(post_check_mixing) <- getMarkerClasses(object)

    cols <- getStockcol()[1:K] # get appropriate colours
    gg <- ggplot(post_check_mixing, aes(x = posteriorShrinkage,
                                        y = posteriorZscore,
                                        color = organelle)) +
        geom_point(size = 5)
    gg <- gg + theme_minimal() +
        theme(panel.grid.major = element_blank(), panel.grid.minor = element_blank(),
              panel.border = element_rect(colour = "black", fill = NA, size = 1),
              plot.title = element_text(hjust = 0.5, size = 20),
              legend.text=element_text(size = 14)) + scale_color_manual(values = cols) +
        ggtitle(label = "") + xlim(c(0, 1))

    return(gg)
}