R/CrossImpute.R

In nbroguiere/burgertools: Collection of tools extending the functionality of Seurat

#' MAGIC imputation of features of an assay based on the distances in another assay
#'
#' This function imputes features in an assay based on (the merger with) another assay. Typically using the high resolution in the RNA data to impute more noisy additional assays.
#'
#' @param object Seurat object.
#' @param impute.assay character(1). The name of the assay which should be imputed, or "metadata" (Default: DefaultAssay(object)).
#' @param name character(1). The name under which the resulting imputed assay is stored in the Seurat object (Default: paste0("imputed_",impute.assay))
#' @param impute.features character(n). Features on which to run the imputation in the impute.assay, or "all" for all features in the impute.assay, or "variable_features" for the variable features in the impute.assay (Default:"all").
#' @param reference.assay character(1). The name of the assay which is appended to the impute.assay to guide imputation distances (Default: "RNA").
#' @param reference.features character(n). Features from the reference.assay to use to guide the imputation of the impute.assay, or "all" for all features in the reference.assay, or "variable_features" for the variable features in the reference.assay (Default:"variable_features").
#' @param npca integer(1). Number of principal components to use for imputation.
#' @param knn integer(1). Number of nearest neighbors on which to compute bandiwth imputation.
#' @param t integer(1). Diffusion parameter passed to MAGIC.
#' @param n.jobs integer(1). Number of threads on which to run the imputation (Default: 6).
#' @param layer.use character(1). Name of the layer to impute (Default: "data").
#' @return Seurat object with an additional layer containing MAGIC cross-imputed data.
#' @keywords MAGIC Rmagic imputation crossimputation cross-imputation
#' @export
#' @examples
#' MySeuratObject <- CrossImpute(MySeuratObject,"other_assay_to_impute_based_on_RNA_distances")
#' MySeuratObject[["imputed-RNA"]] # Check result
CrossImpute <- function(object, impute.assay=DefaultAssay(object), name=paste0("imputed.",impute.assay), impute.features="all", reference.assay="RNA", reference.features="variable_features", npca=40, knn=3, t=2, n.jobs=6, layer.use="data"){

  if(!"Rmagic" %in% rownames(installed.packages())){
    stop('Rmagic is not installed, but needed for imputations. Consider running: devtools::install_github("cran/Rmagic"). The python version magic-impute is also needed, see install instructions at https://github.com/cran/Rmagic.')
  }

  if(!Rmagic::pymagic_is_available()){
    stop("Rmagic is installed, but pymagic is not available. Make sure Rmagic, reticulate, and pymagic are functional. Can be tested with Rmagic::pymagic_is_available().")
  }

  backup.default.assay <- DefaultAssay(object)

  if(impute.assay=="metadata"){
    if(sum(impute.features=="all")){
      warning("Define explicit metadata columns to impute, 'all' is not accepted for metadata imputation.")
      return(object)
    }else if(sum(impute.features=="variable_features")){
      warning("Define explicit metadata columns to impute, 'variable_features' is not accepted for metadata imputation.")
      return(object)
    }else{
      missing.features <- setdiff(unique(unlist(impute.features)), colnames(object@metadata))
      impute.features <- intersect(unique(unlist(impute.features)), colnames(object@metadata))
      if(length(missing.features)>0){
        warning(paste("The following features were not found in the impute.assay of the Seurat object and were omitted: \n",paste(missing.features,collapse=" ")))
      }
    }
    df <- object@metadata[,impute.features,drop=F]
  }else{
    DefaultAssay(object) <- impute.assay
    if(sum(impute.features=="all")){
      impute.features <- rownames(object)
    }else if(sum(impute.features=="variable_features")){
      impute.features <- VariableFeatures(object)
    }else{
      missing.features <- setdiff(unique(unlist(impute.features)), rownames(object))
      impute.features <- intersect(unique(unlist(impute.features)), rownames(object))
      if(length(missing.features)>0){
        warning(paste("The following features were not found in the impute.assay of the Seurat object and were omitted: \n",paste(missing.features,collapse=" ")))
      }
    }
    if(!length(impute.features)){
      warning("No features to impute available, aborting.")
      return(object)
    }else{
      df <- t(as.matrix(GetAssayData(object = object, layer = layer.use)[impute.features,,drop=F]))
    }
  }

  DefaultAssay(object) <- reference.assay
  if(reference.features[1]=="all"){
    reference.features <- rownames(object)
  }else if(reference.features[1]=="variable_features"){
    reference.features <- VariableFeatures(object)
  }

  missing.features <- setdiff(unique(unlist(reference.features)), rownames(object))
  reference.features <- intersect(unique(unlist(reference.features)), rownames(object))
  if(length(missing.features)>0){
    warning(paste("The following features were not found in the reference.assay of the Seurat object and were omitted: \n",paste(missing.features,collapse=" ")))
  }
  if(!length(reference.features)){
    warning("No reference features available, aborting.")
    return(object)
  }else{
    df <- cbind(df,t(as.matrix(GetAssayData(object = object, layer = layer.use)[reference.features,,drop=F])))
  }

  imputed <- Rmagic::magic(data = as.matrix(df), npca=npca, knn=knn, t=t, n.jobs=n.jobs)
  imputed <- t(imputed$result)
  object[[name]] <- CreateAssayObject(imputed[impute.features,])

  DefaultAssay(object) <- backup.default.assay
  return(object)
}