R/TCSScutoff.R

In YuLab-SMU/GOSemSim: GO-terms Semantic Similarity Measures

#' determine the topological cutoff for TCSS method
#'
#' @param OrgDb OrgDb object
#' @param keytype keytype
#' @param ont ontology : "BP", "MF", "CC"
#' @param combine_method "max", "BMA", "avg", "rcmax", "rcmax.avg"
#' @param ppidata A data.frame contains positive set and negative set.
#' Positive set is PPI pairs that already verified.
#' ppidata has three columns, column 1 and 2 are character, column 3
#' must be logical value:TRUE/FALSE.
#'
#' @return numeric, topological cutoff for given parameters
#' @export
#'
#' @examples
#' \dontrun{
#'     library(org.Hs.eg.db)
#'     library(STRINGdb)
#'
#'     string_db <- STRINGdb$new(version = "11.0", species = 9606,
#'     score_threshold = 700)
#'     string_proteins <- string_db$get_proteins()
#'
#'     #get relationship
#'     ppi <- string_db$get_interactions(string_proteins$protein_external_id)
#'
#'     ppi$from <- vapply(ppi$from, function(e)
#'                        strsplit(e, "9606.")[[1]][2], character(1))
#'     ppi$to <- vapply(ppi$to, function(e)
#'                        strsplit(e, "9606.")[[1]][2], character(1))
#'     len <- nrow(ppi)
#'
#'     #select length
#'     s_len <- 100
#'     pos_1 <- sample(len, s_len, replace = T)
#'     #negative set
#'     pos_2 <- sample(len, s_len, replace = T)
#'     pos_3 <- sample(len, s_len, replace = T)
#'     #union as ppidata
#'     ppidata <- data.frame(pro1 = c(ppi$from[pos_1], ppi$from[pos_2]),
#'      pro2 = c(ppi$to[pos_1], ppi$to[pos_3]),
#'      label = c(rep(TRUE, s_len), rep(FALSE, s_len)),
#'      stringsAsFactors = FALSE)
#'
#'     cutoff <- tcss_cutoff(OrgDb = org.Hs.eg.db, keytype = "ENSEMBLPROT",
#'     ont = "BP", combine_method = "max", ppidata)
#' }
tcss_cutoff <- function(OrgDb = NULL, keytype = "ENTREZID", ont,
                        combine_method = "max", ppidata) {

  semdata <- godata(OrgDb, keytype = keytype, ont = ont, computeIC = TRUE,
                    processTCSS = FALSE, cutoff = NULL)
  #cutoff is in the range of ICT value
  IC <- semdata@IC
  GO <- names(IC[!is.infinite(IC)])
  offspring <- switch(ont,
                      MF = AnnotationDbi::as.list(GOMFOFFSPRING),
                      BP = AnnotationDbi::as.list(GOBPOFFSPRING),
                      CC = AnnotationDbi::as.list(GOCCOFFSPRING))
  #compute ICT value for each term
  ICT <- computeICT(GO, offspring)
  #cutoffs, all possible cutoff values
  cutoffs <- seq(0.1, max(ICT) + 0.1, by = 0.1)
  #all genes/proteins that have none-zero annotations
  all_pro <- unique(semdata@geneAnno[, keytype])
  #filter the ppidata
  filtered_ppidata <- create_filtered_ppidata(all_pro, ppidata = ppidata)
  #calculate the similarity value for filtered_ppidata
  predict_result <- lapply(cutoffs, computePre,
                           filtered_ppidata = filtered_ppidata,
                           semdata = semdata,
                           combine_method = combine_method)

  #calculate the auc and F1_score
  auc_F1_score <- calc_auc_F1_score(predict_result,
                                    filtered_ppidata = filtered_ppidata)
  #decide the most appropriate cutoff
  decide_cutoff(auc_F1_score, cutoffs = cutoffs)
}

#' keep the proteins with none-zero annotations
#'
#' @param all_pro all proteins that have none-zero annotations
#' @param ppidata data.frame, already verified PPI data
#' @importFrom stats na.omit
#'
#' @return data.frame, annotated protein pairs and their labels
#' @noRd
create_filtered_ppidata <- function(all_pro, ppidata) {
  #check data type
  if (!(is.character(ppidata[, 1]) && is.character(ppidata[, 2]) &&
        is.logical(ppidata[, 3]))) {
    stop("ppidata must be a data.frame with three columns:character, character, logical")
  }

  ppidata <- na.omit(ppidata)

  #remove proteins that have zero annotations
  len1 <- ppidata[, 1] %in% all_pro
  len2 <- ppidata[, 2] %in% all_pro
  ppidata_exist <- ppidata[len1 & len2, ]
  filtered_ppidata <- unique(ppidata_exist)

  len <- dim(filtered_ppidata)[1]

  if (len == 0) {
    stop("filtered ppidata is empty, none items have GO annotation. Please input more data.")
  }

  nTrue <- sum(filtered_ppidata[, 3])
  nFalse <- len - nTrue

  if (nTrue == len || nFalse == len) {
    stop("The filtered ppidata lacks the necessary label:TRUE and FALSE. Please input more data.")
  }

  message(paste("positive set has", nTrue,
                "PPI pairs, negative set has", nFalse, "PPI pairs"))

  return(filtered_ppidata)
}

#' compute prediction value on filtered ppidata
#'
#' @param cutoff numeric, topological cutoff
#' @param filtered_ppidata data.frame, annotated protein pairs and their labels
#' @param semdata GOSemSimDATA object
#' @param combine_method "max" "BMA", "avg", "rcmax", "rcmax.avg"
#' @return list, the prediction value for the input cutoff
#' @noRd
#'
computePre <- function(cutoff, filtered_ppidata, semdata,
                       combine_method) {
  #tcssdata is updated with this input cutoff
  tcssdata <- process_tcss(semdata@ont, semdata@IC, cutoff = cutoff)

  semdata@tcssdata <- tcssdata
  #similarity value is calculated with the semdata
  mapply(geneSim, MoreArgs = list(semData = semdata,
                                  measure = "TCSS",
                                  combine = combine_method,
                                  drop = FALSE),
         filtered_ppidata[, 1], filtered_ppidata[, 2])
}

#' calculate auc and F1-score
#'
#' @param predict_result list, prediction value for all cutoffs
#' @param filtered_ppidata data.frame, annotated protein pairs and their labels
#' @return data.frame, auc and F1-score value for different cutoffs
#' @importFrom methods slot
#' @importFrom rlang check_installed
#' @noRd
#'
calc_auc_F1_score <- function(predict_result, filtered_ppidata) {
  # checking whether package 'ROCR' being installed
  check_installed('ROCR', 'for`calc_auc_F1_score()`.')
  # the label for PPIs, TRUE/FALSE
  label <- filtered_ppidata[, 3]
  #geneSim returns one value and two characters in once calculation
  value_pos <- seq(from = 1, to = length(label) * 3, by = 3)
  #just the similarity value
  pre_value <- lapply(predict_result, function(p) as.numeric(p[value_pos]))
  #returned value may contains NA
  pos_stay <- !is.na(pre_value[[1]])
  label <- label[pos_stay]

  # prediction object
  pred <- lapply(pre_value, function(e) {
    ROCR::prediction(e[pos_stay], label,
                     label.ordering = c(FALSE, TRUE)
    )
  })
  # performance object for auc
  perf_auc <- lapply(pred, ROCR::performance, measure = "auc")
  # auc value
  auc <- unlist(lapply(perf_auc, function(e) slot(e, "y.values")[[1]]))
  # performance object for F1-score
  perf_F1_score <- lapply(pred, ROCR::performance, measure = "f")
  # F1-score value, average value at different semantic similarity cutoffs
  F1_score <- unlist(lapply(perf_F1_score, function(e) {
    mean(slot(e, "y.values")[[1]], na.rm = TRUE)
  }))

  #return as data.frame
  return(data.frame(auc = auc,
                    F1_score = F1_score,
                    stringsAsFactors = F))
}

#' select the most appropriate cutoff
#'
#' @param auc_F1_score data.frame, auc and F1-score value for different cutoffs
#' @param cutoffs vector, all possible cutoff values
#' @return vector, topological cutoff for given parameters
#' @noRd
#'
decide_cutoff <- function(auc_F1_score, cutoffs) {
  #product value satisfies the "both maximized" requirement
  auc_mutiply_F1 <- auc_F1_score[, "auc"] * auc_F1_score[, "F1_score"]
  #get the max product value
  pos <- which(auc_mutiply_F1 == max(auc_mutiply_F1))

  if (length(pos) == 1)  return(cutoffs[pos])

  #if not only one pair of auc and F1-score have same product
  #take the one with larger auc
  select_auc <- auc_F1_score[pos, "auc"]

  auc_pos <- which(select_auc == max(select_auc))

  if (length(auc_pos) == 1) return(cutoffs[pos[auc_pos]])

  #if more than one pair of auc and F1-score are both same
  #take the smaller cutoff for time saving
  return(cutoffs[pos[min(auc_pos)]])
}