Leave-One-Out-Cross-Validation Based Machine Learning Algorithms for Efficient, Scalable, and Robust Prediction of Artemisinin Resistance in Plasmodium falciparum Kelch13 mutants.

Abstract

Emerging Plasmodium falciparum resistance to frontline artemisinin (ART) derivatives threatens global malaria control efforts. Current protocols assigning ART-sensitive (ART-S) or -resistant (ART-R) phenotypes to parasite strains with Kelch13 (K13) mutations remain time-intensive for real-time surveillance. Ring-survival assay (RSA), an ex vivo gold standard mandates laboratory-adaptation under culture conditions and frequently does not replicate clinical ART-responsiveness, thereby limiting its utility for rapidly triaging newly emerging K13 variants. In contrast, advent of artificial intelligence (AI)-derived models hold potential to screen >100 K13 variants within hours, enabling rapid near-real-time prediction of ART-S/R. Here, we engineered a panel of forty representative K13 variants by computing their relative deviations from wild-type K13 across ten physicochemical and structural descriptors, yielding a compact feature set which captures mutation-induced perturbations in folding, stability, and solvent exposure. These can be further utilized to train and validate AI classifier to predict resistance.