R/compareModels.R

In MOFA: Multi-Omics Factor Analysis (MOFA)

################################################
## Functions to compare different MOFA models ##
################################################


#' @title Correlation of the latent factors across different trials
#' @name compareFactors
#' @description Different objects of \code{\link{MOFAmodel}} are compared in terms of correlation between 
#' their latent factors. The correlation is calculated only on those samples which are present in all models.
#' Ideally, the output should look like a block diagonal matrix, suggesting that
#'  all detected factors are robust under different initializations.
#' If not, it suggests that some factors are weak and not captured by all models.
#' @param models a list containing \code{\link{MOFAmodel}} objects.
#' @param comparison type of comparison, either 'pairwise', i.e. compare one model
#'  with another one at a time, or 'all', i.e. calculate correlation between factors from all model.
#'  By default, all models are compared.
#' @param show_rownames logical indicating whether to show rownames in heatmap (see also pheatmap documentation)
#' @param show_colnames logical indicating whether to show colnames in heatmap (see also pheatmap documentation)
#' @param ... extra arguments passed to pheatmap
#' @details This function can be helpful to evaluate the robustness of factors across
#'  different random initilizations. 
#' Large block of factors from different models in the correlation matrix show consistent factors,
#'  while stand-alone factors that are only recovered in a single model instance are less reliable.
#' @return Plots a heatmap of correlation of Latent Factors in all models when 'comparison' is 'all'. 
#' Otherwise, for each pair of models, a seperate heatmap is produced comparing one model againt the other.
#' The corresponding correlation matrix or list or pairwise correlation matrices is returned
#' @importFrom stats cor
#' @importFrom stats var
#' @importFrom pheatmap pheatmap
#' @importFrom grDevices colorRampPalette
#' @export
#' @examples
#' ### Example on simulated data
#' # Simulate Data
#' data <- makeExampleData()
#' # Create MOFA model
#' MOFAobject <- createMOFAobject(data)
#' # Prepare MOFA model
#' MOFAobject <- prepareMOFA(MOFAobject)
#' # Train several instances of MOFA models
#' n_inits <- 3 
#' MOFAlist <- lapply(seq_len(n_inits), function(i) runMOFA(MOFAobject, outfile=tempfile()))
#' compareFactors(MOFAlist, comparison="all")
#' compareFactors(MOFAlist, comparison="pairwise")


compareFactors <- function(models, comparison = "all", show_rownames=FALSE, show_colnames=FALSE,...) {
  
  # Sanity checks
  if(!is.list(models))
    stop("'models' has to be a list")
  if (!all(vapply(models, function (l) is(l, "MOFAmodel"), logical(1))))
    stop("Each element of the the list 'models' has to be an instance of MOFAmodel")
  if (!comparison %in% c("all", "pairwise"))
    stop("'comparison' has to be either 'all' or 'pairwise'")
  
  # give generic names if no names present
  if(is.null(names(models))) names(models) <- paste("model", seq_along(models), sep="")
  
  # get latent factors
  LFs <- lapply(seq_along(models), function(modelidx){
    model <- models[[modelidx]]
    Z <- getFactors(model)
    Z
    })
  
  for (i in seq_along(LFs)) 
    colnames(LFs[[i]]) <- paste(names(models)[i], colnames(LFs[[i]]), sep="_")
  
  if (comparison=="all") {
    #get common samples between models
    commonSamples <- Reduce(intersect,lapply(LFs, rownames))
    if(is.null(commonSamples)) 
      stop("No common samples in all models for comparison")
    
    #subset LFs to common samples
    LFscommon <- Reduce(cbind, lapply(LFs, function(Z) {
      Z <- Z[commonSamples,, drop=FALSE]
      nonconst <- apply(Z,2,var, na.rm=TRUE) > 0
      if(sum(nonconst) < ncol(Z)) message("Removing ", sum(!nonconst), " constant factors from the comparison.")
      Z[, nonconst]
    })
      )

    # calculate correlation
    corLFs <- cor(LFscommon, use="complete.obs")
    
    #annotation by model
    # modelAnnot <- data.frame(model = rep(names(models), times=vapply(LFs, ncol, numeric(1))))
    # rownames(modelAnnot) <- colnames(LFscommon)
    
    #plot heatmap
    # if(is.null(main)) main <- "Absolute correlation between latent factors"
    if(length(unique(as.numeric(abs(corLFs))))>1){
    pheatmap(abs(corLFs), show_rownames = show_rownames,show_colnames = show_colnames,
             color = colorRampPalette(c("white",RColorBrewer::brewer.pal(9,name="YlOrRd")))(100),
             # color=colorRampPalette(c("white", "orange" ,"red"))(100), 
             # annotation_col = modelAnnot, main= main , ...)
             ...)
    } else warning("No plot produced as correlations consist of only one value")
    # return(corLFs)
  }
  
  if(comparison=="pairwise"){
    PairWiseCor <- lapply(seq_along(LFs[-length(LFs)]), function(i){
      LFs1<-LFs[[i]]
        sublist <- lapply((i+1):length(LFs), function(j){
          LFs2<-LFs[[j]]
          common_pairwise <- intersect(rownames(LFs1), rownames(LFs2))
          if(is.null(common_pairwise)) {
            warning(paste("No common samples between models",i,"and",j,"- No comparison possible"))
            NA
          }
          else{
          # if(is.null(main)) main <- paste("Absolute correlation between factors in model", i,"and",j)
          nonconst1 <- apply(LFs1[common_pairwise,],2,var, na.rm=TRUE) > 0
          nonconst2 <- apply(LFs2[common_pairwise,],2,var, na.rm=TRUE) > 0
          if(sum(!nonconst1) + sum(!nonconst2)>0) 
            message("Removing ", sum(!nonconst1), " and " ,sum(!nonconst2),
                    " constant factors from the each model for the comparison, respectively.")
          corLFs_pairs <- cor(LFs1[common_pairwise,nonconst1],
                              LFs2[common_pairwise,nonconst2],
                              use="complete.obs")
          if(length(unique(abs(corLFs_pairs)))>1){
          pheatmap(abs(corLFs_pairs),color=colorRampPalette(c("white", "orange" ,"red"))(100), ...)
          } else warning("No plot produced as correlations consist of only one value")
          corLFs_pairs
          }
        })
        names(sublist) <- names(models)[(i+1):length(LFs)]
        sublist
    })
    
    names(PairWiseCor) <- names(models[-length(models)])
    return(NULL)
    #return(PairWiseCor)
  }
}


#' @title Compare different instances of trained \code{\link{MOFAmodel}} 
#' @name compareModels
#' @description Different objects of \code{\link{MOFAmodel}} are compared in terms
#'  of the final value of the ELBO statistics. 
#' For model selection the model with the highest ELBO value is selected.
#'  The height of the bar indicates the number of inferred factors and 
#'  the color of the bar the value of the ELBO statistic.
#' @param models a list containing \code{\link{MOFAmodel}} objects.
#' @param show_modelnames boolean, whether to indicate the name of each model instance
#'  (names of the list in models) or not
#' @return a ggplot showing the number of factors and
#'   the ELBO statistics of the given models as a barplot
#' @export
#' @examples
#' ### Example on simulated data
#' # Simulate Data
#' data <- makeExampleData()
#' # Create MOFA model
#' MOFAobject <- createMOFAobject(data)
#' # Prepare MOFA model
#' MOFAobject <- prepareMOFA(MOFAobject)
#' # Train several instances of MOFA models
#' n_inits <- 3 
#' MOFAlist <- lapply(seq_len(n_inits), function(i) runMOFA(MOFAobject, outfile=tempfile()))
#' compareModels(MOFAlist)

compareModels <- function(models, show_modelnames = FALSE) {
  # Sanity checks
  if(!is.list(models))
    stop("'models' has to be a list")
  if (!all(vapply(models, function (l) is(l,"MOFAmodel"), logical(1))))
    stop("Each element of the the list 'models' has to be an instance of MOFAmodel")

  elbo_vals <- vapply(models, getELBO, numeric(1))
  n_factors <- vapply(models, function(m) {
    n_fac <- getDimensions(m)$K
    n_fac
    }, numeric(1))
  if(is.null(names(models))) names(models) <- paste0("model_", seq_along(models))
  df <- data.frame(ELBO=elbo_vals, model = names(models), n_factors=n_factors)
  gg <- ggplot(df, aes_string(x="model",y="n_factors", fill="ELBO")) + geom_bar(stat="identity")+
    ylab("Number of inferred factors")
  if(show_modelnames) gg <- gg + theme(axis.text.x=element_text(angle=60, vjust=1, hjust=1))
  else gg <- gg + theme(axis.text.x=element_blank())
  return(gg)
}

#' @title Select the best model from a list of trained \code{\link{MOFAmodel}} objects
#' @name selectModel
#' @description Different trained objects of \code{\link{MOFAmodel}} are compared
#'  in terms of the final value of the ELBO statistics 
#' and the model with the highest ELBO value is selected.
#' @param models a list containing \code{\link{MOFAmodel}} objects.
#' @param plotit show a plot of the characteristics of the compared
#'   \code{\link{MOFAmodel}} objects (ELBO value and number of inferred factors)?
#' @return a single \code{\link{MOFAmodel}} with the best ELBO statistics from the provided list 
#' @export
#' @examples
#' # Simulate Data
#' data <- makeExampleData()
#' # Create MOFA model
#' MOFAobject <- createMOFAobject(data)
#' # Prepare MOFA model
#' MOFAobject <- prepareMOFA(MOFAobject)
#' # Train several instances of MOFA models
#' n_inits <- 3 
#' MOFAlist <- lapply(seq_len(n_inits), function(i) runMOFA(MOFAobject, outfile=tempfile()))
#' selectModel(MOFAlist)

selectModel <- function(models, plotit =TRUE) {
  # Sanity checks
  if(!is.list(models))
    stop("'models' has to be a list")
  if (!all(vapply(models, function (l) is(l, "MOFAmodel"), logical(1))))
    stop("Each element of the the list 'models' has to be an instance of MOFAmodel")

  elbo_vals <- vapply(models, getELBO, numeric(1))
  if(plotit) print(compareModels(models))
  models[[which.max(elbo_vals)]]
}