example

Determination of HIV-1 Coreceptor Usage via the PART rule generator

Below is a condensed version of the rules that give predictions of coreceptor usage when PART is selected. Based on cross-validation experiments, these rules are expected to classify unseen test cases with 89% accuracy. They classifies 250 out of our 271 training cases correctly (92.25%). Here is a breakdown of PART's performance in cross-validation.

TP Rate	FP Rate	Precision	Recall	Class
0.777	0.042	0.92	0.777	cxcr4
0.958	0.223	0.875	0.958	ccr5

PART decision list
------------------

p12 = S AND
p35 = I: ccr5 (156.0/14.0)

p12 = R: cxcr4 (69.0/2.0)

p8 = T AND
p6 = N AND
p16 = - AND
p23 = F: ccr5 (21.0/1.0)

: cxcr4 (25.0/4.0)

If translated into english, this would read "If there is a S at position 12 and an I at position 35 then ccr5, otherwise if there is a R at position 12 then cxcr4, otherwise if there is a T an position 8, a N at position 6, a gap at position 16, and an F at position 23 then ccr5. In all other cases cxcr4". Notice that rules are read in order, so that if there is a conflict, the higher rule wins.

The numbers in parantheses to the right of the rules indicate how many samples from the training set are covered by that rule versus how many were incorrectly classified. (# covered / # misclassified) .