Generative adversarial neural network with graph embeddings for de novo designing small-molecule inhibitors against Mycobacterium tuberculosis KasA enzyme

A generative semi-supervised adversarial neural network trained on graph embeddings was developed for de novo design of potential inhibitors against beta-ketoacyl-[acyl-carrier protein] synthase I (KasA), an enzyme critically important for biosynthesis of mycolic acids of the Mycobacterium tuberculosis cell wall. The designed model was trained and tested on a set of compounds from a virtual library of small molecules containing structural elements capable of selective interactions with the therapeutic target. Using the developed neural network, 3,637 compounds were de novo designed, followed by assessment of their inhibitory activity against the KasA protein using molecular docking methods. Based on the analysis of the obtained data, six compounds exhibiting high affinity to the malonyl-binding site of the enzyme were selected. The identified compounds are assumed to form promising basic structures for further theoretical and experimental studies on the development of new effective inhibitors of drug-resistant tuberculosis.