The use of machine learning to improve the effectiveness of ANRS in predicting HIV drug resistance.

Abstract

BACKGROUD HIV has placed a large burden of disease in developing countries. HIV drug resistance is inevitable due to selective pressure. Computer algorithms have been proven to help in determining optimal treatment for HIV drug resistance patients. One such algorithm is the ANRS gold standard interpretation algorithm developed by the French National Agency for AIDS Research AC11 Resistance group. OBJECTIVES The aim of this study is to investigate the possibility of improving the accuracy of the ANRS gold standard in predicting HIV drug resistance. METHODS Data consisting of genome sequence and a HIV drug resistance measure was obtained from the Stanford HIV database. Machine learning factor analysis was performed to determine sequence positions where mutations lead to drug resistance. Sequence positions not found in ANRS were added to the ANRS rules and accuracy was recalculated. RESULTS The machine learning algorithm did find sequence positions, not associated with ANRS, but the model suggests they are important in the prediction of HIV drug resistance. Preliminary results show that for IDV 10 sequence positions where found that were not associated with ANRS rules, 4 for LPV, and 8 for NFV. For NFV, ANRS misclassified 74 resistant profiles as being susceptible to the ARV. Sixty eight of the 74 sequences (92%) were classified as resistance with the inclusion of the eight new sequence positions. No change was found for LPV and a 78% improvement was associated with IDV. CONCLUSION The study shows that there is a possibility of improving ANRS accuracy.