Abstract submitted to Retroviruses and Opportunistic Infection Conference

10/14/02

Background: The use of resistance testing for study entry into the TORO-1 and TORO-2 phase III clinical trials has made available a large comprehensive dataset of reverse transcriptase (RT) and protease (PR) genotypes from highly HAART-experienced patients (pts) in different geographical regions for analysis in a predictive decision tree model.

Methods: Pt genotype and phenotype data were collected using the Geneseq and PhenoSense at ViroLogic, Inc. The complete genotype dataset for RT and PR (including polymorphic variations from reference) of pts (n = 491) from the U.S., Canada, Mexico, and Brazil (TORO-1) were used to construct the tree models. Tree-modeling methods (CHAID and CART methods) generated a decision tree using a randomly assigned learning set of samples (n = 396) and verified by a testing data set (n = 95). The resulting decision trees for both reverse transcriptase and protease inhibitors were further evaluated with the second phase III study population (n = 499) from Europe and Australia (TORO-2).

Results: Using the methods described above, predictive decision trees could be established for all 15 antiretrovirals currently approved. The final trees contained 5–11 nodes with well characterized resistance mutation positions in RT or PR as the junction points. The majority of trees formed a predictive pathway containing less than 8 nodes. The models led to correct prediction rates in the range of 78.4%–94.9% for reverse transcriptase inhibitors (nucleoside and non-nucleoside classes) and 82.6%–92.0% for protease inhibitors for the TORO-1 samples. The TORO-2 data set showed correct prediction rates ranging from 77.0%–93.6% for reverse transcriptase inhibitors and from 83.4%–89.0% for protease inhibitors.

Conclusions: The results of the predictive tree-modeling method used here were comparable to, or superior to, recently published analyses. For some antiretrovirals such as lamivudine, nevirapine, and indinavir, the trees built from viral genotypes of TORO-1 pts robustly predicted the response for samples from TORO-2. This work demonstrated the possibility of generating a predictive tree model from a large, globally diverse heavily-treated study population of HIV-1 pts.