Background:  HIV-1 rapidly accumulates mutations in reverse transcriptase (RT) in response to selection by RT inhibitors (RTI). At least 24 codon sites in RT have been implicated in resistance to RTI, with increasing evidence of distinct genetic pathways of resistance. However, comparative methods for associating variation in RTI resistance with variation in RT sequences among patients have failed to account for founder effects that may cause false associations. To address this issue, we carry out an analysis of within-host polymorphisms in RT as revealed by mixtures (i.e. ambiguous nucleotides) in the context of evolving resistance to RTI using Bayesian networks, which can identify patterns of statistical associations characteristic of causal relationships.

Methods:  We obtained 6571 HIV-1 subtype B sequences for the first 300 codons of RT from the Stanford HIV Resistance database, corresponding to 3560 patients administered regimens that utilized 11 different RTI and 3011 RTI-naive patients (screened for RTI-resistant variants). To evaluate the effect of selection for RTI resistance among hosts, we carried out a selection analysis for each patient group using HyPhy. We identified 21,288 non-synonymous mixtures and analyzed the joint distribution of mixtures and RTI using Bayesian networks in HyPhy, in order to identify associations between the emergence of polymorphisms and specific RTI.

Results:  Sites under significant positive selection were highly concordant among patient groups, with 9 additional sites in the RTI-treated group. Substitutions at 5 of the additional sites (L74, K103, M184, Y188, T215) had known associations with RTI resistance, whereas the remaining 4 did not (T69, R83, A98, Q207). Our network identified 22 sites conditionally dependent on specific RTI. (Constraining RTI to depend on sites reduced the posterior probability of the network by at least 100-fold.) We also identified co-variation among sites, some that defined pathways of resistance, e.g., T215 was conditionally dependent on both atazanavir and M41; D67 on atazanavir and K70; and V108 on efavirenz and H221.

Conclusions:  We resolved a detailed map of the evolution of RTI resistance in HIV-1 populations within hosts from sequencing mixtures. This method is free from founder effects and provides a probabilistic model for predicting RTI-specific resistance from RT sequences.