A Multi-Property Optimizing Generative Adversarial Network for de novo Antimicrobial Peptide Design

Antimicrobial peptides (AMPs) play a crucial role in developing novel anti-infective drugs due to their broad-spectrum antimicrobial activity and lower likelihood of causing bacterial resistance. However, laboratory synthesis of AMPs is tedious and time-consuming. Existing computational methods have limited capability in optimizing multiple desired properties simultaneously. Here, a Multi-Property Optimizing Generative Adversarial Network (MPOGAN) is proposed to iteratively learn the relationship between peptides and multi-properties with a dynamically updated dataset. With the increase of the dataset quality, the ability of the model to design AMPs with multiple desired properties is improved. Through extensive computational tests, MPOGAN exhibits superior performance in generating AMPs with multiple desired properties, including potent antimicrobial activity, reduced cytotoxicity, and increased diversity. Ten designed AMPs are chemically synthesized, nine of which exhibited antimicrobial activity and low cytotoxicity. Notably, two of these peptides showed potent broad-spectrum antimicrobial activity coupled with reduced cytotoxicity, highlighting their potential for downstream applications.