ScholarWorks@성균관대학교

초록

The World Health Organization warns that antimicrobial resistance (AMR) poses a global threat to public health and calls for the discovery of new antimicrobial peptides (AMPs) as potential therapeutic alternatives. Artificial intelligence (AI) revolutionized the search for AMPs, yet existing generative models like those based on GANs or diffusion often lack systematic evaluation of trade-offs between fidelity, novelty, and diversity. Moreover, sequence length is rarely treated as an explicit controllable design parameter. Here, we propose CoLPAT-AMP, a Transformer-based expert system enabling partial control over sequence length in AMP generation. Two complementary models were implemented and compared: (1) a conditioned model, which integrates ESM-2 embeddings with physicochemical property and sequence length conditioning during training (without explicit property specification at inference), and (2) an unconditioned model, trained solely on AMP sequences. The conditioned model achieved lower reconstruction loss (0.99 vs. 1.73), higher reconstruction accuracy (0.70 vs. 0.53), novelty (1 vs. 0.85), and greater diversity. While the overlap between physicochemical properties of generated sequences and real AMPs was stronger for the unconditioned model, indicating higher adherence to the training dataset distribution. An independent AMP classifier further confirmed that the main portion of the generated sequences by both models maintained antimicrobial characteristics. These findings suggest that the conditioned model enables partial control and expands exploration toward more diversity, while the unconditioned model emphasizes property fidelity. Overall, CoLPAT-AMP provides a property-aware, and partially length-controllable framework for rational AMP design, advancing AI-driven strategies to explore peptide space and combat antimicrobial resistance.

키워드