Self-imitation learning for AMP design (SIL-AMP)

The growing threat of antimicrobial resistance highlights the need for efficient computational methods to discover novel antimicrobial peptides (AMPs). However, navigating the vast combinatorial sequence space while balancing exploration and optimization remains a major challenge. We introduce SIL-AMP, a self-imitation learning optimization approach for AMP design that iteratively refines a transformer-based generative policy using its own top evaluated sequences by the objective function. Per optimization round, SIL-AMP employs incremental stochastic beam search to generate diverse peptide candidates, which are evaluated using APEX Pathogen model to predict their minimum inhibitory concentrations (MICs) against clinically relevant bacterial pathogens. A greedy selection procedure identifies high-scoring candidates while enforcing validity constraints. Rather than relying on explicit value-function estimation, SIL-AMP fine-tunes its generative policy through next-token cross-entropy learning on the action trajectories of its top-performing candidates, reinforcing the actions that lead to promising peptide designs in subsequent iterations. By alternating between diverse sequence generation, objective function-guided evaluation, and self-imitation learning, SIL-AMP provides an iterative approach for exploring and optimizing AMP sequence space, enabling the prioritization of novel peptide candidates with favorable predicted antimicrobial activity for subsequent experimental validation.