DARWIN: A Framework for Target Specific Diversity Constrained Natural Product Like Molecule Generation

The rising global incidence of incurable diseases underscores the persistent gaps in drug discovery and development, largely rooted in the limited chemical diversity of modern drug compounds. Natural products (NP)—chemical metabolites produced by living organisms—offer a rich reservoir of diversity to address this limitation. Therefore, this study develops a novel framework, DARWIN, a genetic-algorithm based framework, that leverages the diversity of NPs and the scalability of computational techniques to propose novel Natural Product like drug candidates. While genetic algorithms have been widely used in molecule optimization, the molecules generated by them are known to lack diversity. DARWIN supports fine-grained control over molecular diversity by incorporating intermolecular similarity directly within the generation process. Since they are based on genetic algorithms, they are extremely efficient without the need for expensive pretraining on GPUs, or finetuning for targeted generation. When applied to two targets implicated in Ewing Sarcoma and Chronic Lymphocytic Leukemia, the generated molecules demonstrated improved properties relative to the DOCKSTRING baseline. Overall, DARWIN provides a novel, controllable framework for expanding the search for drug candidates beyond synthetic libraries, offering an effective method for accelerating drug discovery for currently incurable diseases.