<tool id="ROC-PR Curve" name="ROC-PR Curve">
	<description>calculates AUC for ROC and PR curves</description>
	<command interpreter="sh">r_wrapper.sh $script_file</command>
	<inputs>
		<param format="tabular" name="cvpred_data" type="data" label="CV Predictions"/>
	</inputs>
	<outputs>
		<!--
			<data format="pdf" name="rocprc.pdf" from_work_dir="rocprc.pdf" label="ROC-PR Curve" />
		-->
		<data format="png" name="rocprc.png" from_work_dir="rocprc.png" />
	</outputs>

	<configfiles>
		<configfile name="script_file">

			rm(list = objects() )

			########## calculate auprc #########
			auPRC &lt;- function (perf) {
				rec &lt;- perf@x.values
				prec &lt;- perf@y.values
				result &lt;- list()
				for (i in 1:length(perf@x.values)) {
				result[i] &lt;- list(sum((rec[[i]][2:length(rec[[i]])] - rec[[i]][2:length(rec[[i]])-1])*prec[[i]][-1]))
				}
				return(result)
			}

			########## plot ROC and PR-Curve #########
			rocprc &lt;- function(x) {
				sink(NULL,type="message")
				options(warn=-1)
				suppressMessages(suppressWarnings(library('ROCR')))
				svmresult &lt;- data.frame(x)
				colnames(svmresult) &lt;- c("Seqid","Pred","Label", "CV")

				linewd &lt;- 1
				wd &lt;- 4
				ht &lt;- 4
				fig.nrows &lt;- 1 
				fig.ncols &lt;- 2
				pt &lt;- 10
				cex.general &lt;- 1 
				cex.lab &lt;- 0.9
				cex.axis &lt;- 0.9
				cex.main &lt;- 1.2
				cex.legend &lt;- 0.8


				#pdf("rocprc.pdf", width=wd*fig.ncols, height=ht*fig.nrows)
				png("rocprc.png", width=wd*fig.ncols, height=ht*fig.nrows, unit="in", res=100)

				par(xaxs="i", yaxs="i", mar=c(3.5,3.5,2,2)+0.1, mgp=c(2,0.8,0), mfrow=c(fig.nrows, fig.ncols))

				CVs &lt;- unique(svmresult[["CV"]])
				preds &lt;- list()
				labs &lt;- list()
				auc &lt;- c()
				for(i in 1:length(CVs)) {
					preds[i] &lt;- subset(svmresult, CV==(i-1), select=c(Pred))
					labs[i] &lt;- subset(svmresult, CV==(i-1), select=c(Label))
				}

				pred &lt;- prediction(preds, labs)
				perf_roc &lt;- performance(pred, 'tpr', 'fpr')
				perf_prc &lt;- performance(pred, 'prec', 'rec')

				perf_auc &lt;- performance(pred, 'auc')
				prcs &lt;- auPRC(perf_prc)
				avgauc &lt;- 0
				avgprc &lt;- 0

				for(j in 1:length(CVs)) {
					avgauc &lt;- avgauc + perf_auc@y.values[[j]]
					avgprc &lt;- avgprc + prcs[[j]]
				}        

				avgauc &lt;- avgauc/length(CVs)
				avgprc &lt;- avgprc/length(CVs)

				#preds_merged &lt;- unlist(preds)
				#labs_merged &lt;- unlist(labs)
				#pred_merged &lt;- prediction(preds_merged, labs_merged)
				#perf_merged_auc &lt;- performance(pred_merged, 'auc')

				plot(perf_roc, colorize=T, main="ROC curve", spread.estimate="stderror",
				xlab="1-Specificity", ylab="Sensitivity", cex.lab=1.2)
				text(0.2, 0.1, paste("AUC=", format(avgauc, digits=3, nsmall=3)))

				plot(perf_prc, colorize=T, main="P-R curve", spread.estimate="stderror",
				xlab="Recall", ylab="Precision", cex.lab=1.2, xlim=c(0,1), ylim=c(0,1))
				text(0.2, 0.1, paste("AUC=", format(avgprc, digits=3, nsmall=3)))

				dev.off()
			}

			############## main function #################
			d &lt;- read.table("${cvpred_data}")

			rocprc(d)

		</configfile>
	</configfiles>

	<help>
	
**Note**

This tool is based on the ROCR library.  If you use this tool please cite:
		
Tobias Sing, Oliver Sander, Niko Beerenwinkel, Thomas Lengauer.
ROCR: visualizing classifier performance in R.
Bioinformatics 21(20):3940-3941 (2005).  
        
----

**What it does**

Takes as input cross-validation predictions and calculates ROC Curve and its area under curve (AUC) and PR Curve and its AUC.

----

**Results**

ROC Curve: Receiver Operating Characteristic Curve. Compares true positive rate (sensitivity) to false positive rate (1 - specificity).  

PR Curve: Precision Recall Curve. Compares number of true positives (recall; same as sensitivity) to the number of true positives relative to the total number sequences classified as positive (precision). 

AUC for a given curve: Area Under the Curve: Probability that of a randomly selected positive/negative pair, the positive will be scored more highly by the trained SVM than a negative.

.. class:: infomark

Both curves measure SVM performance, but ROC curves can be inaccurate if there is a large skew in class distribution.  For more information see:

Jesse Davis, Mark Goadrich.
The Relationship Between Precision-Recall and ROC Curves.
Proceedings of the 23rd Annual Internation Conference on Machine Learning.
Pittsburgh, PA, 2006.

----

**Example**

.. image:: ./static/images/sample_roc_chen.png
	</help>
</tool>