Background:  Despite extensive experience using genotypic and phenotypic data for HIV drug resistance interpretation, evolution of drug resistance in vivo remains difficult to predict. Here we model combinations of drugs as causal but imperfect influences on viral mutation. The resulting “noisy-OR” relation is closely related to a log-linear dependence of mutation probabilities on drug combinations, allowing it to be trained directly from clinical records.

Methods:  We extract “genetic selection episodes” from clinical records based on 2 genotypic assays of resistance to therapy, 1 at the beginning of a new drug therapy and 1 after treatment has given rise to new mutations. Logistic regression is used over a system of 33 NRTI or NNRTI treatment-selected mutations, and related to 9 reverse transcriptase (RT) inhibitors (6 NRTI and 3 NNRTI). For training and testing, 558 genetic selection episodes extracted from 6837 patients’ clinical data were used. Cross-validation (10x) and a threshold on model deviance identifies 19 statistically significant RT mutations for analysis.

Results:  In general, the correspondence between drug/mutation coefficients established by regression from genetic selection episodes with HIV database treatment-selected mutation “scores” (based on genotypic and phenotypic data) is striking: from 171 (19 mutations x 9 RT inhibitors) potential interactions, 29 statistically significant (p <0.05) coefficients correspond in 25 cases to significant (≥5) HIV database scores, including a strong, negative hyper-sensitive coefficient for mutation L74:AZT. E44:d4T is the only significant positive model coefficient with a lower HIV database score, and it has been mentioned by others. Next, we go beyond conventional intra-RTI-class effects to consider all RTI as potential causes of all RT mutations. Here too the regression coefficients generally confirm known NRTI/NRTI-related- treatment-selected mutations and NNRTI/NNRTI-related- treatment-selected mutations diagonal interactions. However, the model also predicts 7 other, significant off-diagonal drug:mutation interactions:  NNRTI-induced, NRTI-associated:  K70:NVP, L74:EFV, L210:EFV, K219:DLV; and NRTI-induced, NNRTI-associated:  V108:D4T, V179:3TC, V179:d4T.

Conclusions:  A straight-forward, general logistic regression technique can be used with clinical records containing genotypic testing and multi-drug combination therapy data. Results recapitulate many drug:mutation interactions known from anecdotal, sequence or phenotypic-testing, systematizes them, and predict several new NRTI/NNRTI cross-class mutations.