Predicting Dihydroartemisinin Resistance in Plasmodium falciparum using Pathway Activity Inference

Drug resistance threatens the effectiveness of treatments of infectious diseases, particularly on the global scale where mutation is rapid, mechanisms of resistance are developing or unknown, and limited data is available. Pathway activity inference is a dimensionality reduction method with proven effectiveness in classifying cancer types and drug responses based on transcription data. We propose a novel application of pathway activity inference to predict dihydroartemisinin resistance in the Plasmodium falciparum strain of malaria, a global infectious disease. Optimized pathway activity inference models outperform untransformed gene expression models in both in vitro regression (p = 0.03) and in vivo classification tasks (p = 2 × 10− 9). Optimal methods were found to be mostly ensemble (5 of 12) and/or kernel-based (7 of 12), providing the first evidence of the effectiveness of kernel methods for predicting drug resistance in infectious diseases. Performance metrics of the optimal in vitro model on in vivo data (accuracy , area under receiver operating characteristic curve = 0.63) affirmed the low empirical correlation between resistance measures in the two settings.