Protein language models predict artemisinin partial resistance and prioritize genetic modifiers in Plasmodium falciparum

Authors:

Abstract
Artemisinin partial resistance threatens malaria control, but molecular surveillance relies largely on established K13 variants, limiting interpretation of emerging sequence variation. Here we used ESM-3 protein representations to model resistance in 1,293 Plasmodium falciparum clinical isolates. K13 sequence models generalized across temporally independent cohorts (auROC > 0.88; auPR > 0.83) and retained predictive information for geographically distinct and previously unseen K13 substitutions, correctly classifying 13 of 17 variants absent from training. Extending the framework to 167 ART-R-associated candidate genes recovered established resistance loci and provided complementary predictive information to training-selected single-variant markers. Model-based prioritization identified EXO E415G, and CRISPR–Cas9 editing showed that E415G increased ring-stage survival only in parasites carrying K13 C580Y, whereas ATP4 G1128R had no measurable effect. These findings establish protein-sequence modeling as a complementary approach for prioritizing emerging resistance-associated variation and context-dependent genetic modifiers.

This publication has 55 references indexed in Scilit: