MultiPepGen: A conditional generative framework for the de novo design of function-specific antimicrobial peptides

The accelerating global crisis of antimicrobial resistance (AMR) has intensified the need for innovative therapeutic strategies. Although antimicrobial peptides (AMPs) represent a promising alternative to conventional antibiotics, their experimental discovery remains costly and time-intensive. Computational approaches have accelerated AMP candidate discovery; however, the ability to specifically target distinct biological activities remains a critical challenge for optimizing downstream experimental validation. Addressing this need, we introduce MultiPepGen, a conditional generative framework integrating adversarial learning with recurrent sequence modeling under explicit functional conditioning. Its architectural design is systematically examined through ablation analysis across recurrent variants. A multi-criteria evaluation framework combining distributional fidelity, statistical descriptor analysis, and external classifier-based functional prediction indicates that MultiPepGen preserves key distributional characteristics of natural AMPs while balancing sequence diversity with predicted functional enrichment. Crucially, our analysis reveals a potency-diversity trade-off: models selected for high classifier confidence exhibit a measurable contraction of the physicochemical descriptor space, whereas models favoring stable generative dynamics preserve broader exploratory capacity. Downstream candidate filtering identified top de novo candidates exhibiting high predicted potency and less than 45% sequence identity to the training AMP set, indicating substantial sequence-level novelty. By providing both an open-source implementation and an intuitive graphical interface, MultiPepGen offers an accessible computational pipeline for targeted antimicrobial peptide design.