Prediction of phenotypic susceptibility to antiretroviral drugs using physiochemical properties of the primary enzymatic structure combined with artificial neural networks

Objectives: Genotypic interpretation systems extrapolate observed associations in datasets to predict viral susceptibility to antiretroviral drugs (ARVs) for given isolates. We aimed to develop and validate an approach using artificial neural networks (ANNs) that employ descriptors of physiochemical properties for mutations in HIV-1 protease (PR) and reverse transcriptase (RT) to predict phenotypic susceptibility to all currently approved ARVs. Method: We extracted pairs of PR and RT gene sequences (n = 1507; 98.5% sub-type B) and their corresponding exact phenotype values (PhenoSense only, n = 10132) from the Stanford HIV database. All amino acid positions and mixture codes were accounted for. For each ARV, an ANN was trained with 10-fold internal cross-validation. The predictive abilities of these trained ANNs were validated on separate datasets. Results: Correlation coefficients between observed and predicted phenotype values in the 10-fold cross-validation ranged from: 0.75 (tenofovir) to 0.94 [lamivudine (3TC)] for nucleoside RT inhibitors (NRTIs); 0.82 [efavirenz (EFV)] to 0.83 [nevirapine (NVP)] for non-nucleoside RT inhibitors (NNRTIs); and 0.83 (atazanavir) to 0.92 (ritonavir) for PR inhibitors (PIs). For the validation set the correlation coefficients ranged from 0.76 (didanosine) to 0.96 (3TC) for NRTIs; 0.68 (EFV) to 0.81 (NVP) for NNRTIs; and 0.88 (amprenavir) to 0.95 (saquinavir) for PIs. For C sub-type predictions, with ANNs trained on sub-type B data, the correlation coefficient was 0.89. Conclusions: ANNs, based on the physiochemical properties of the PR and RT amino-acid sequences, predict phenotypic susceptibility to ARVs inhibiting these enzymes to an extent that is comparable to routine phenotypic susceptibility testing. These ANNs can also be used to predict resistance to C sub-types.