first implementation of domain-specific F1

metric/domain-specific-f1/domain-specific-f1.r

@@ -0,0 +1,111 @@
+#!/usr/bin/env Rscript
+
+# Author_and_contribution: Niklas Mueller-Boetticher; created template
+# Author_and_contribution: Mark D. Robinson; coded the domain-specific F1
+
+suppressPackageStartupMessages(library(optparse))
+
+# TODO adjust description
+option_list <- list(
+  make_option(
+    c("-l", "--labels"),
+    type = "character", default = NULL,
+    help = "Labels from domain clustering."
+  ),
+  make_option(
+    c("-g", "--ground_truth"),
+    type = "character", default = NA,
+    help = "Groundtruth labels."
+  ),
+  make_option(
+    c("-e", "--embedding"),
+    type = "character", default = NA,
+    help = "Embedding of points in latent space. Potential usage for metrics without groundtruth."
+  ),
+  # format should be json
+  make_option(
+    c("-c", "--config"),
+    type = "character", default = NA,
+    help = "Optional config file (json) used to pass additional parameters."
+  ),
+  make_option(
+    c("-o", "--out_file"),
+    type = "character", default = NULL,
+    help = "Output file."
+  )
+)
+
+# TODO adjust description
+description <- "Calculate domain-specific F1 score (returns JSON with vector: F1 for each true domain)"
+
+opt_parser <- OptionParser(
+  usage = description,
+  option_list = option_list
+)
+opt <- parse_args(opt_parser)
+
+# Use these filepaths as input
+label_file <- opt$labels
+
+if (!is.na(opt$ground_truth)) {
+  groundtruth_file <- opt$ground_truth
+}
+if (!is.na(opt$embedding)) {
+  embedding_file <- opt$embedding
+}
+if (!is.na(opt$config)) {
+  config_file <- opt$config
+}
+
+
+## Code for calculating metric goes here
+## --------------------------------------
+library(clue)
+library(jsonlite)
+
+# # for testing - start
+# label_file <- "results/libd_dlpfc/Br8100_151675/BayesSpace/domains.tsv"
+# groundtruth_file <- "data/libd_dlpfc/Br8100_151675/labels.tsv"
+# outfile <- "./domain-specific-f1.json"
+# # for testing - stop
+
+domains <- read.delim(label_file, sep="\t", row.names = 1)
+groundtruth <- read.delim(groundtruth_file, sep="\t", row.names = 1)
+
+rn <- intersect(rownames(domains), rownames(groundtruth))
+
+# subset to common set
+domains <- domains[rn,,drop = FALSE]
+groundtruth <- groundtruth[rn,,drop = FALSE]
+
+(tb <- table(domains$label, groundtruth$label))
+
+hg <- clue::solve_LSAP(tb, maximum = TRUE)
+sa <- seq_along(hg)
+mapping <- data.frame(cluster = rownames(tb)[sa], 
+                      label = colnames(tb)[hg],
+                      tp = tb[cbind(seq_along(hg), hg)])
+rownames(mapping) <- mapping$label
+
+tt <- table(groundtruth$label)
+mapping[names(tt),"domain_size"] <- tt
+
+tl <- table(domains$label)
+m <- match(mapping$cluster, names(tl))
+mapping$cluster_size <- as.numeric(tl[m])
+
+mapping$recall <- with(mapping, tp/domain_size)
+mapping$precision <- with(mapping, tp/cluster_size)
+mapping$f1 <- with(mapping, 2/(1/recall+1/precision))
+
+# rownames(mapping) <- NULL
+
+## Write output
+outfile <- file(opt$out_file)
+write_json(unname(mapping[,c("label","f1")]),outfile)
+
+# from template
+# dir.create(dirname(outfile), showWarnings = FALSE, recursive = TRUE)
+# 
+# writeLines(format(metric, digits = 6, scientific = TRUE), outfile)
+# close(outfile)

metric/domain-specific-f1/domain-specific-f1.yml

@@ -0,0 +1,8 @@
+channels:
+  - conda-forge
+dependencies:
+  - r-base=4.3.1
+  - r-mclust=6.0.0
+  - r-clue=0.3_65
+  - r-optparse=1.7.3
+  - r-jsonlite=1.8.7