Machine learning-based generative models have enabled the efficient and accurate exploration of vast chemical space in recent decades. However, many approaches do not explicitly account for the synthetic feasibility of generated molecules due to the challenge of integrating both theoretical and experimental perspectives. To overcome this challenge, here we propose a reaction-regulated graph-based genetic algorithm, namely R2GB-GA, that enables synthetically feasible molecular editing through effectively exploring complex chemical space for molecular design. By embedding domain-specific reaction rules into the evolutionary process, our method enables site-selective molecular modifications to a scaffold that allows for both preserving specific scaffolds and changing only specific molecular substructures. We show that the proposed method generates more synthetically accessible molecules across diverse generative tasks, outperforming conventional approaches when evaluated using fragment-based and pathway-based scoring methods. In addition, we illustrate the practical applicability of our method to drug design through ligand design targeting the inhibition of heat shock protein 90, a representative chaperone protein in cancer therapy. We believe that with the use of a reaction-based framework, our approach could be applied as a general method for a broad range of drug discovery strategies including late-stage functionalization.
Moon et al. (Sun,) studied this question.