Controllable Generation of Pathogen-Specific Antimicrobial Peptides Through Knowledge-Aware Prompt Diffusion Model

Generative models have shown considerable promise in antimicrobial peptide design; however, their ability to generate pathogen-specific peptides remains limited due to data scarcity. In this study, KPPepGen, a controllable generative framework that leverages knowledge-aware pathogen prompts derived from pre-training on Gene Ontology and pathogen knowledge graphs, which act as knowledge injections to guide a diffusion model in generating biologically plausible peptides tailored to specific pathogens, is introduced. Then, KPPepGen is extended for peptide optimization by integrating prompt-guided partial diffusion with multi-site combinatorial mutations. Experimental results show that KPPepGen can simultaneously generate valid peptides for 56 distinct pathogens, achieving high novelty, favorable physicochemical properties, and delivering over a 10% performance improvement for pathogens with limited training data. Further analysis demonstrates that KPPepGen effectively captures essential sequence and structure patterns characteristic of individual pathogens. The optimization results reveal a high success rate of 44.3% for Magainin 2, along with an average improvement of 7.6% compared to the ESM-based method, underscoring the effectiveness of KPPepGen in enhancing the overall performance of peptides. Finally, for clinically relevant pathogens such as E. coli and S. aureus, KPPepGen successfully generated nine novel peptides that exhibit strong antimicrobial activity and low cytotoxicity in the wet-lab evaluation.