E-mail Abstract Author Session Search Abstracts Program

Session 71 Poster Presentations
Resistance to HIV-1 Protease Inhibitors
Session Day and Time: Tuesday 1:30 - 3:30 pm
Room: Hall A


594
Genotypic Basis of Variation in Protease Susceptibility in Primary HIV Infection Analyzed with Machine Learning
AJ Leigh Brown*1,2, SDW Frost2, B Good2, ES Daar3, V Simon4, M Markowitz4, AC Collier5, E Connick6, B Conway7, JB Margolick8, JP Routy9, J Corbeil2, N Hellmann10, DD Richman2, SJ Little2
1Univ of Edinburgh, Scotland; 2Univ of California at San Diego; 3Harbor-UCLA, Los Angeles, CA; 4Aaron Diamond AIDS Res Ctr, New York, NY; 5Univ of Washington, Seattle; 6Univ of Colorado, Denver; 7Univ of British Columbia, Vancouver, Canada; 8Johns Hopkins Univ, Baltimore, MD; 9McGill Univ Hlth Ctr, Montreal, Canada; and 10ViroLogic, San Francisco, CA

Background: As many as 20% of individuals with primary HIV infection have a virus strain that is hyper susceptible to amprenavir or ritonavir; many of these strains also show reduced replicative capacity (RC). In antiretroviral (ARV)-treated patients (pts), hyper susceptibility (HS) to amprenavir has been associated with N88S in PR, but this mutation is not seen in untreated pts. Other studies have suggested that variation in gag contributes to HS. The aim of this study was to use machine learning to analyze the genotypic basis of this variation in phenotype.

Methods: Samples were obtained from individuals in acute or early HIV infection who had not received any ARV therapy. Phenotypes were determined using HIV PhenoSense. Genotypes were determined on an ABI automated sequencer. Hypersusceptibility was taken as an IC50 0.4-fold or less of the IC50 of HIV-1NL4-3. The J4.8 implementation of the C4.5 decision tree algorithm was used to generate models which were cross-validated by 90/10 splits of the data.

Results: The sequence dataset (n = 188) included all polymorphic amino acid (aa) sites and insertions from p7gag through PR, a total of 94 variables. In the PR-based version, only the 20 polymorphic sites from PR itself were included. Both datasets were used to construct a model for HS to ritonavir. Overall % correct classifications for PR- and gag + PR-based models were 72% and 73%, respectively. For PR alone, sensitivity (true positive HS) = 73%, specificity = 68%; for gag + PR data, sensitivity = 59%, specificity = 75%. The structure of the PR model was: (PR57 (PR10 (PR63 (PR37 (PR13)(PR62 (PR93)))))), while that from gag + PR was (PR57 (G418 (G473 (G486 (PR61 (G474 (PR63))))))): no insertions were included in the model. Of 22 ritonavir HS cases, 10 were predicted by both models, 3 only by the gag + PR model and 5 only by the PR model. There was a strong association between predicted HS and low RC values (£ 10%): 7/10 cases of low RC were included in the HS cases identified by one or both models, confirming the close relationship between these two phenotypes.

Conclusions: Variation in PI susceptibility in primary HIV infection is not due to single mutations but to combinations of amino acids at several polymorphic sites. Several pathways to ritonavir HS have been identified. Incorporating such models into clinical studies will allow the value of baseline genotype in predicting pts’ responses to PIs to be assessed.