roc: ROC for NEA benchmarks

Description

Plot ROC curve(s) for benchmarked network(s)

Usage

roc(tpvsfp, mode = "roc", tpr.stop = 1, fpr.stop = 1, coff.z = 1.965,
  coff.fdr = 0.1, cex.main = 1, cex.leg = 0.75, main = NA,
  use.lty = FALSE, print.stats = TRUE, sort_by_letter_and_remove = FALSE)

Arguments

tpvsfp

a list contaning data for the ROC curve(s) from benchmark: cross.z, cutoffs, ne, nv,tp, fp, tn, fn.

mode

either ROC curve (roc) or precision recall curve (prc). default: roc

tpr.stop

upper limit for the Y-axis (true positives). default: 1

fpr.stop

upper limit for the X-axis (false prositives), default: 1

coff.z

the point where to stop the ROC curve (in order to disable, set coff.z to below the minimal possible level)

coff.fdr

the point where to put the circle ('cross') at the ROC curve

cex.main

font size for the main title

cex.leg

font size for the legend

main

title name for the plot, default: none

use.lty

if different line types should be used for the curves, useful when the number of curves is very big (>10). default:"FALSE"

print.stats

add N(edges) and N(vertices) to the legend lines; only works if non-empty $ne and $nv elements are submitted in tpvsfp. default:"TRUE"

sort_by_letter_and_remove

enable a specific order for the curves in the legend, set by prefix letters gsub("[A-Z]\:", "", p1, fixed=F, ignore.case=T), such as "a:network1", "b:network-2" etc., whereby a: and b: are removed from the text strings. default:"FALSE"

Details

The function roc can be called either from inside benchmark, or separately. In the latter case, it should receive the first argument as an object of the same type as benchmark can return. Generally, a ROC curve encompasses the whole interval from 0% to 100% at both axes. However in case of predictions made via network analysis, only the interval with (at least) formally significant scores is of interest (Merid et.al 2014). Therefore the benchmark results are presented as ROC curves that end at minimal acceptable confidence, set via either a z-score cut-off. Additionally, an extra point is denoted that might correspond to cross.z in roc. The scale is given in no. of tested member genes rather then as percentage of the total no. of tests. The scale ranges xlim and ylim of the plot can be reduced by submitting parameters tpr.stop and fpr.stop with values <1.

References

http://www.biomedcentral.com/1471-2105/15/308

Examples

Run this code

# NOT RUN {
# Benchmark and plot one networks on the whole set of test GSs, using no mask:
data(can.sig.go);
fpath <- can.sig.go
gs.list <- import.gs(fpath, Lowercase = 1, col.gene = 2, col.set = 3);
data(net.kegg)
netpath <- net.kegg
net <- import.net(netpath)
# }
# NOT RUN {
b0 <- benchmark (NET = net,
 GS = gs.list,
 echo=1, graph=TRUE, na.replace = 0, mask = ".", minN = 0,
 coff.z = 1.965, coff.fdr = 0.1, Parallelize=2);
 roc(b0, coff.z = 1.64);
# }
# NOT RUN {
## Benchmark and plot a number of networks on GO terms and KEGG pathways separately, using masks
b1 <- NULL;
for (mask in c("kegg_", "go_")) {
b1[[mask]] <- NULL;
for (file.net in c("netpath")) {
# a series of networks can be put here: c("netpath1", "netpath2", "netpath3")
net <- import.net(netpath, col.1 = 1, col.2 = 2, Lowercase = 1, echo = 1)
b1[[mask]][[file.net]] <- benchmark (NET = net, GS = gs.list,
gs.gene.col = 2, gs.group.col = 3, net.gene1.col = 1, net.gene2.col = 2,
echo=1, graph=FALSE, na.replace = 0, mask = mask, minN = 0, Parallelize=2);
}}
par(mfrow=c(2,1));
roc(b1[["kegg_"]], coff.z = 2.57,main="kegg_");
roc(b1[["go_"]], coff.z = 2.57,main="go_");
# }
# NOT RUN {
# }