PIRSpred: a web server for reliable HIV-1 protein-inhibitor resistance/susceptibility prediction.

We introduce a publicly available webserver, PIRSpred (http://protinfo.compbio.washington.edu/pirspred/), for accurate human immunodeficiency virus type 1 (HIV-1) genotypic resistance/susceptibility prediction. The server accepts mutant HIV-1 protease or reverse transcriptase (RT) enzyme sequences as input and predicts resistance/ susceptibility to several FDA-approved inhibitors using three approaches: linear regression, docking with dynamics, and their consensus. The predictions made by the server outperform other publicly available HIV-1 genotypic interpretation algorithms, and can be used efficiently and economically in a clinical setting to guide treatment against HIV-1 infection. The development of antiretroviral drug resistance severely impedes treatment of HIV-1-infected patients. The International AIDS Society-USA [1] has recommended that antiretroviral drug regimens be adjusted for patientswho fail to suppressHIV-1viral load to!200 copies per ml. Genotypic testing provides viral mutation information that might suggest new drug regimens to suppress the growth of drug-resistant variants [2]. However, interpreting genotypic results can be problematic owing to complex mutational patterns [3]. To overcome this problem, several groups have developed knowledge-based techniques, including linear regression, neural networks, decision trees, support vector machines and rule-based methods, to interpret HIV-1 genotypic data (refer to Ref. [4] and references therein). These methods generally perform well, but need large viral phenotype-genotype training sets and often make discordant phenotypic predictions [5]. The accuracies are particularly low for newly approved drugs because there are fewer genotype-phenotype results available for learning. Physics-based approaches that evaluate the binding energy of protein–inhibitor complexes have successfully identified key mutations in parts of the enzyme that confer drug resistance and show a high correlation with the inhibitory concentration (IC50) values determinedbyphenotypic tests (refer to Ref. [6] and references therein). This approach achieves higher accuracy than the knowledgebased approaches for newly approved drugs because it does not require a large phenotype-genotype database for learning, but is limited when considering mutations that are not located at the active site where the drug binds. Here, we introduce the webserver PIRSpred, which uses both knowledgeand physics-based approaches, to