Machine Learning connects structure, bitterness & mechanism to antimalarial activity

Abstract

Machine learning allows us to identify patterns we might otherwise have missed in the data. It therefore provides an ample solution to the age-old problem of lossless research. When the pattern is robust and the signal is durable, it is possible to use relatively modest amounts of data to build powerful well-learned models. By constructing a classifier using gradient-boosted machines, it was shown that the model was robust, yielded high quality metrics and demonstrated consistent probability assignments for chemistries that have a suitable chemical backbone qualified for inclusion in antiplasmodic libraries. Critical to model development was the utilization of molecular fingerprinting and extraction of physico-chemical parameters relevant to the mechanism driving antiplasmodic activity. Subsequently, such an approach allows the model to uncover the link between bitterness, molecular structure and therapeutic value. This approach to evaluating antiplasmodic activity in chemistries provides a low-cost tool usable in identifying new classes of molecules for use in reducing malaria morbidity that often affects vulnerable members of the community the most. Additionally, given their relatively broad, low colligation and potent efficacies, these molecules may provide strong safety margins and durability netting high returns for health equity.