Calculating ROC curves

(Original entry published in CBDD Research Group Blog.)

Here you will find a a short tutorial about how to generate receiver operating characteristic (ROC) curves and other statistics after running RxDock molecular docking (for other programs such as Vina or Glide, just a little modification on the way dataforR_uq.txt file is interpreted will make it work, see below).

I assume all of you are familiar with what ROC curves are, what are they for and how they are made.

Just in case, a very brief summary would be:

  • ROC curves are graphic representations of the relation existing between the sensibility and the specificity of a test. It is generated by plotting the fraction of true positives out of the total actual positives versus the fraction of false positives out of the total actual negatives.

  • In our case, we will use it for checking whether a docking program is able to select active ligands with respect to inactive ligands (decoys) and whether it is able to select these active ligands in the top % of a ranked database.

  • R Library ROCR is mandatory (try with command install.packages("ROCR") in R before downloading from source).

The example selected for this tutorial is a system from the DUD benchmark set, “hivpr” or “hiv protease”.

These are the files you will need (all can be downloaded in this Dropbox shared folder):

  • List of active ligands (ligands.txt)

  • List of inactive ligands (decoys.txt)

  • Output file with the docked poses of each ligand with the corresponding docking scores (hivpr_all_results.sd.gz)

  • R script with all the R commands in this tutorial (ROC_curves.R)

Before getting into R, the resulted docked poses have to be filtered out for only having the best pose for each ligand (the smallest score – or highest in negative value). To do so run:

gunzip hivpr_all_results.sd.gz
sdsort -n -s -fSCORE hivpr_all_results.sd | sdfilter -f'$_COUNT == 1' > hivpr_1poseperlig.sd
# sdsort with -n and -s flags will sort internally each ligand by increasing
# score and sdfilter will get only the first entry of each ligand.

sdreport -t hivpr_1poseperlig.sd | awk '{print $2,$3,$4,$5,$6,$7}' > dataforR_uq.txt
# sdreport will print all the scores of the output in a tabular format and,
# with command awk, we will format the results.

Note

sdsort and sdreport are really useful tools for managing sd formatted compound collections. They are very user-friendly and free to download. They are provided along with rDock software in the Download section of the website.

This dataforR_uq.txt (also in the Dropbox folder) file must contain one entry per ligand with the docked scores (what R will use to rank and plot the ROC curves).

R commands for generating ROC curves

Then, run the following commands in R for plotting the ROC curves:

# load ROCR
library(ROCR);

# load ligands and decoys
lig <- unique(read.table("ligands.txt")[,1]);
dec <- unique(read.table("decoys.txt")[,1]);

# load data file from docking
uniqRes <- read.table("dataforR_uq.txt", header=T);

# change colnames
colnames(uniqRes)[1]="LigandName";

# add column with ligand/decoy info
uniqRes$IsActive <- as.numeric(uniqRes$LigandName %in% lig)

# define ROC parameters
# here INTER is selected to compare between ligands using SCORE.INTER
# this could be changed for also running with other programs
predINTERuq <- prediction(uniqRes$INTER*-1, uniqRes$IsActive)
perfINTERuq <- performance(predINTERuq, 'tpr', 'fpr')

# plot in jpg format with a grey line with theoretical random results
jpeg("hivpr_Rinter_ROC.jpg")
plot(perfINTERuq, main="hivpr - ROC Curves", col="blue")
abline(0, 1, col="grey")
dev.off()

Which will give us the following plot:

../_images/hivpr_Rinter_ROC.jpg

Afterwards, other useful statistics such as AUC or Enrichment factors can also be calculated:

# AUC (area under the curve)
auc_rdock <- performance(predINTERuq, "auc")
auc.area_rdock <- slot(auc_rdock, "y.values")[[1]]
cat("AUC: \n")
cat(auc.area_rdock)
cat("\n\n")
AUC:
0.7700965
# Enrichment Factors
EF_rdock <- perfINTERuq@y.values[[1]] / perfINTERuq@x.values[[1]]
EF_rdock_1 <- EF_rdock[which(perfINTERuq@x.values[[1]] > 0.01)[1]]
EF_rdock_20 <- EF_rdock[which(perfINTERuq@x.values[[1]] > 0.2)[1]]
cat("Enrichment Factor top 1%:\n")
cat(EF_rdock_1)
cat("\n\n")
Enrichment Factor top 1%:
11.11817
cat("Enrichment Factor top 20%:\n")
cat(EF_rdock_20)
cat("\n\n")
Enrichment Factor top 20%:
3.200686

Moreover, a good analysis of these curves is to re-plot them in semilogarithmic scale (x axis in logarithmic scale). This way, one can focus on the early enrichment of the database and have a more detailed view of the selected actives in the top % of all the ligands.

jpeg("hivpr_semilog_ROC.jpg")
rdockforsemilog=perfINTERuq@x.values[[1]]
rdockforsemilog[rdockforsemilog < 0.0005]=0.0005
plot(rdockforsemilog, perfINTERuq@y.values[[1]],type="l", xlab="False Positive Rate", ylab="True Positive Rate", xaxt="n", log="x", col="blue", main="hivpr - Semilog ROC Curves")
axis(1, c(0, 0.001, 0.01, 0.1, 1))
x<-seq(0, 1, 0.001)
points(x, x, col="gray", type="l")
dev.off()

Obtaining the following semi-logarithmic ROC curves:

../_images/hivpr_semilog_ROC.jpg