Machine learning prioritizes synthesis of primaquine ureidoamides with high antimalarial activity and attenuated cytotoxicity

Primaquine (PQ) is a commonly used drug that can prevent the transmission of Plasmodium falciparum malaria, however toxicity limits its use. We prepared five groups of PQ derivatives: amides 1a-k, ureas 2a-k, semicarbazides 3a,b, acylsemicarbazides 4a-k and bis-ureas 5a-v, and evaluated them for antimalarial activity in vitro against the erythrocytic stage of P. falciparum NF54. Particular substituents, such as trityl (in 2j and 5r) and methoxybenzhydryl (in 3b and 5v) were associated with a favorable cytotoxicity-to-activity ratio. To systematically link structural features of PQ derivatives to antiplasmodial activity, we performed a quantitative structure-activity relationship (QSAR) study using the Support Vector Machines machine learning method. This yielded a highly accurate statistical model (R2 = 0.776 in cross-validation), which was used to prioritize novel candidate compounds. Seven novel PQ-ureidoamides 10a-g were synthesized and evaluated for activity, highlighting the benzhydryl ureidoamides 10e and 10f derived from p-chlorophenylglycine. Further experiments on human cell lines revealed that 10e and 10f are an order of magnitude less toxic than PQ in vitro while having antimalarial activity indistinguishable from PQ. The toxicity profile of novel compounds 10 toward human cells was particularly favorable when the glucose-6-phosphate dehydrogenase (G6PD) was inhibited, while toxicity of PQ was exacerbated by G6PD inhibition. Our work therefore highlights promising lead compounds for the development of effective antimalarial drugs that may also be safer for G6PD-deficient patients. In addition, we provide computational inferences of antimalarial activity and cytotoxicity for thousands of PQ-like molecular structures.