Semi- Supervised Learning for Predicting Anti-HIV Drug Resistance

An infection with HIV is one of the most challenging infectious disease worldwide. Despite the large amount of anti-retroviral drugs, the therapy success is limited. This limitation results from the fact that the virus genome is integrated into the host’s genome and from the development of drug resistance. Resistance testing is an important tool in therapy management to improve the success of anti-retroviral therapy. There are two ways for testing: Phenotyping, where the resistance factor is measured directly by experiments and genotyping, where important regions of the genome are scanned for mutations. There are a few useful predictive tools (e. g. Geno2Pheno), that try to predict resistance from genotypic data by applying supervised statistical methods. These methods need a lot of labeled data which is expensive and time-consuming to obtain. The present work evaluates the application of Semi-Supervised Learning (SSL) to improve the prediction of resistance from the viral genome if only a small number of labeled samples are available. The results yield that SSL has the capabilities to improve the prediction, especially if the number of labeled samples is small compared to the amount of unlabeled data. Although, not all methods are helpful for all drug classes. Furthermore, the experiments yield that also unlabeled data of the same drug class as the drug in question can help to improve the prediction with SSL.