The Application of Artificial Neural Networks for Phenotypic Drug Resistance Prediction: Evaluation and Comparison with Other Interpretation Systems

Abstract

Although phenotypic resistance testing provides more direct measurement of antiretroviral drug resistance than genotypic testing, it is costly and time-consuming. However, genotypic resistance testing has the advantages of being simpler and more accessible, and it might be possible to use the data obtained for predicting quantitative drug susceptibility to interpret complex mutation combinations. This study applied the Artificial Neural Network (ANN) system to predict the HIV-1 resistance phenotype from the genotype. A total of 7,598 pairs of HIV-1 sequences, with their corresponding phenotypic fold change values for 14 antiretroviral drugs, were trained, validated, and tested in ANN modeling. The results were compared with the HIV-SEQ and Geno2pheno interpretation systems. The prediction performance of the ANN models was measured by 10-fold cross-validation. The results indicated that by using the ANN, with an associated set of amino acid positions known to influence drug resistance for individual antiretroviral drugs, drug resistance was accurately predicted and generalized for individual HIV-1 subtypes. Therefore, high correlation with the experimental phenotype may help physicians choose optimal therapeutic regimens that might be an option, or supporting system, of FDA-approved genotypic resistance testing in heavily treatment-experienced patients.