R/machinelearning-functions-tagm-mcmc-helper.R

In pRoloc: A unifying bioinformatics framework for spatial proteomics

##' Helper function to get the number of outlier at each MCMC
##' iteration.
##'
##' @title Number of outlier at each iteration of MCMC
##' @param x Object of class `MCMCParams`
##' @return A `list` of length `length(x)`.
##' @rdname mcmc-helpers
##' @md
mcmc_get_outliers <- function(x) {
    stopifnot(inherits(x, "MCMCParams"))
    lapply(x@chains@chains, function(mc) coda::mcmc(colSums(1 - mc@Outlier)))
}


##' Helper function to get mean component allocation at each MCMC
##' iteration.
##' @title Mean component allocation at each MCMC iteration
##' @return A `list` of length `length(x)`.
##' @rdname mcmc-helpers
##' @md
mcmc_get_meanComponent <- function(x) {
  stopifnot(inherits(x, "MCMCParams"))
  lapply(x@chains@chains, function(mc) coda::mcmc(colMeans(mc@Component)))
}


##' Helper function to get mean probability of belonging to outlier at
##' each iteration.
##' @title Mean outlier probability
##' @return A `list` of length `length(x)`.
##' @rdname mcmc-helpers
##' @md
mcmc_get_meanoutliersProb <- function(x) {
  stopifnot(inherits(x, "MCMCParams"))
  lapply(x@chains@chains, function(mc) coda::mcmc(colMeans(mc@OutlierProb[, ,2])))
}


##' Wrapper for the geweke diagnostics from coda package also return p-values.
##' @title Geweke diagnostics
##' @param k A `list` of [coda::mcmc] objects, as returned by
##'     `mcmc_get_outliers`, `mcmc_get_meanComponent` and
##'     `mcmc_get_meanoutliersProb`.
##' @return A `matrix` with the test z- and p-values for each chain.
##' @rdname mcmc-helpers
##' @md
geweke_test <- function(k) {
    res <- matrix(NA, nrow = 2, ncol = length(k))
    gwk <- sapply(k, coda::geweke.diag, simplify = TRUE)
    res[1, ] <- unlist(gwk[1, ])
    res[2, ] <- pnorm(abs(unlist(gwk[1, ])), lower.tail = FALSE) * 2
    colnames(res) <- paste0("chain ", seq.int(k))
    rownames(res) <- c("z.value", "p.value")
    return(res)
}


##' Helper function to pool chains together after processing
##'
##' @title Pool MCMC chains
##' @param param An object of class `MCMCParams`.
##' @return A pooled `MCMCParams` object.
##' @rdname mcmc-helpers
##' @md
mcmc_pool_chains <- function(param) {
  stopifnot(inherits(param, "MCMCParams"))

  param1 <- chains(param)[[1]]

  n <- param1@n
  nPool <- param1@n * length(param) # total number of iteration increase
  KPool <- param1@K  # number of components unchanged
  NPool <- param1@N  # number of proteins doesn't change
  numChains <- length(param)

  pooled.Component <- matrix(0, nrow = NPool, ncol = nPool)
  pooled.ComponentProb <- array(0, c(NPool, nPool, KPool ))
  pooled.Outlier <- matrix(0, nrow = NPool, ncol = nPool)
  pooled.OutlierProb <- array(0, c(NPool, nPool, 2 ))

  rownames(pooled.Component) <- rownames(param1@Component)
  rownames(pooled.ComponentProb) <- rownames(param1@ComponentProb)
  rownames(pooled.Outlier) <- rownames(param1@Outlier)
  rownames(pooled.OutlierProb) <- rownames(param1@OutlierProb)
  dimnames(pooled.ComponentProb)[[3]] <- dimnames(param1@ComponentProb)[[3]]


  ## Calculate basic quantities
  for (j in seq_len(numChains)) {

    mc <- chains(param)[[j]]
    ## Pool chains
    pooled.Component[, n * (j - 1) + seq.int(n)] <- mc@Component
    pooled.ComponentProb[, n * (j - 1) + seq.int(n), ] <- mc@ComponentProb
    pooled.Outlier[, n * (j - 1)+ seq.int(n)] <- mc@Outlier
    pooled.OutlierProb[, n * (j - 1) + seq.int(n), ] <- mc@OutlierProb

  }

  mk.list <- lapply(param@chains@chains,function(x) x@ComponentParam@mk)
  lambdak.list <- lapply(param@chains@chains,function(x) x@ComponentParam@lambdak)
  nuk.list <- lapply(param@chains@chains,function(x) x@ComponentParam@nuk)
  sk.list <- lapply(param@chains@chains,function(x) x@ComponentParam@sk)

  ## save Component parameters
  .ComponentParam <- .ComponentParam(K = KPool, D = param1@ComponentParam@D,
                                     mk = Reduce("+", mk.list) / length(mk.list),
                                     lambdak = Reduce("+", lambdak.list) / length(lambdak.list),
                                     nuk = Reduce("+", nuk.list) / length(nuk.list),
                                     sk = Reduce("+", sk.list) / length(sk.list))
  ## apply thinning and burn-in
  .Component <- pooled.Component
  .ComponentProb <- pooled.ComponentProb
  .Outlier <- pooled.Outlier
  .OutlierProb <- pooled.OutlierProb

  ## make MCMCChain object
  .MCMCChain <- .MCMCChain(n = nPool,
                           K = KPool,
                           N = NPool,
                           Component = .Component,
                           ComponentProb = .ComponentProb,
                           Outlier = .Outlier,
                           OutlierProb = .OutlierProb,
                           ComponentParam = .ComponentParam)

  ## Make MCMCChains with single object
  .MCMCChains <- .MCMCChains(chains = list(.MCMCChain))

  ## Make MCMCParams object
  .MCMCParams(method = "TAGM.MCMC",
              chains = .MCMCChains,
              priors = param@priors,
              summary = .MCMCSummary())

}



##' Helper function to burn n iterations from the front of the chains
##'
##' @title MCMC chain burning
##' @param n `integer(1)` defining number of iterations to burn. The default is
##' `50`
##' @return An updated `MCMCParams` object.
##' @rdname mcmc-helpers
##' @md
mcmc_burn_chains <- function(x, n = 50) {
    stopifnot(inherits(x, "MCMCParams"))
    n <- as.integer(n[1])
    stopifnot(is.numeric(n))
    .chain <- chains(x)[[1]]
    K <- .chain@K # Number of components
    N <- .chain@N # Number of Proteins
    chainlist <-
        lapply(x@chains@chains, function(chain) {
            .ComponentParam <- chain@ComponentParam
            ## Subset MCMC iterations
            retain <- seq.int(n + 1, chain@n) ## samples to retain
            ## Check correct number of iterations
            stopifnot(ncol(chain@Component[, retain]) == (chain@n - n))
            .Component <- chain@Component[, retain]
            .ComponentProb <- chain@ComponentProb[, retain, ]
            .Outlier <- chain@Outlier[, retain]
            .OutlierProb <- chain@OutlierProb[, retain, ]
            .MCMCChain(n = as.integer(chain@n - n),
                       K = K,
                       N = N,
                       Component = .Component,
                       ComponentProb = .ComponentProb,
                       Outlier = .Outlier,
                       OutlierProb = .OutlierProb,
                       ComponentParam = .ComponentParam)
        })

    mcmc_chainlist <- .MCMCChains(chains = chainlist)
    .MCMCParams(method = "TAGM.MCMC",
                chains = mcmc_chainlist,
                priors = x@priors,
                summary = .MCMCSummary())
}


##' Helper function to subsample the chains, known informally as
##' thinning.
##'
##' @title MCMC chain thinning
##' @param freq Thinning frequency. The function retains every `freq`th iteration
##' and is an `integer(1)`. The default thinning frequency is `5`.
##' @return A thinned `MCMCParams` object.
##' @rdname mcmc-helpers
##' @author Laurent Gatto
mcmc_thin_chains <- function(x, freq = 5) {
  stopifnot(inherits(x, "MCMCParams"))
  .chain <- chains(x)[[1]]
  K <- .chain@K # Number of components
  N <- .chain@N # Number of Proteins
  nThin <- floor(.chain@n/freq) # Number of iterations after thinning
  chainlist <-
    lapply(x@chains@chains, function(chain) {
      .ComponentParam <- chain@ComponentParam
      ## Subset MCMC iterations
      retain <- freq * seq.int(1:nThin) ## samples to retain
      ## Check correct number of iterations
      stopifnot(ncol(chain@Component[, retain]) == nThin)
      .Component <- chain@Component[, retain]
      .ComponentProb <- chain@ComponentProb[, retain, ]
      .Outlier <- chain@Outlier[, retain]
      .OutlierProb <- chain@OutlierProb[, retain, ]
      .MCMCChain(n = as.integer(nThin),
                 K = K,
                 N = N,
                 Component = .Component,
                 ComponentProb = .ComponentProb,
                 Outlier = .Outlier,
                 OutlierProb = .OutlierProb,
                 ComponentParam = .ComponentParam)
    })

  mcmc_chainlist <- .MCMCChains(chains = chainlist)
  .MCMCParams(method = "TAGM.MCMC",
              chains = mcmc_chainlist,
              priors = x@priors,
              summary = .MCMCSummary())
}



##' Produces a violin plot with the protein posterior probabilities
##' distributions for all organelles.
##'
##' @title Plot posterior probabilities
##' @param y A `character(1)` with a protein name.
##' @param ... Currently ignored.
##' @return A ggplot2 object.
##' @rdname mcmc-helpers
##' @rdname mcmc-plot
setMethod("plot", c("MCMCParams", "character"),
          function(x, y, ...) {
              mcmc_plot_probs(x, y, n = 1)
          })

## Plotting function for violins using ggplot2.
mcmc_plot_probs <- function(param, fname, n = 1) {
    Organelle <- Probability <- NULL
    stopifnot(inherits(param, "MCMCParams"))
    stopifnot(length(fname) == 1)
    chain <- chains(param)[[n]]
    stopifnot(fname %in% rownames(chain@ComponentProb))
    dfr <- as.data.frame(chain@ComponentProb[fname, , ])
    colnames(dfr) <- rownames(chain@ComponentParam@mk)
    dfr_long <- data.frame(Organelle = rep(names(dfr), each = nrow(dfr)),
                           Probability = unlist(dfr, use.names = FALSE),
                           row.names = NULL,
                           stringsAsFactors = FALSE)
    gg2 <- ggplot(dfr_long,
                  aes(Organelle, Probability,
                      width = (Probability))) +
        geom_violin(aes(fill = Organelle), scale = "width", bw = 0.05)
    gg2 <- gg2 + theme_bw() +
        scale_fill_manual(values = pRoloc::getStockcol()[seq_len(nrow(dfr))]) +
        theme(axis.text.x = element_text(angle = 90, hjust = 1),
              axis.title.x = element_blank())
    gg2 <- gg2 +
        ylab("Membership Probability") +
        ggtitle(paste0("Distribution of Subcellular Membership for Protein ", fname ))
    gg2 <- gg2 +
        theme(legend.position = "none")
    return(gg2)
}