Controllable generation of predicted non-hemolytic antimicrobial peptides by multi-guided latent diffusion

Abstract

Antimicrobial peptides (AMPs) are attractive anti-infective scaffolds, but their development is often constrained by hemolysis and by the narrow physicochemical separation between bacterial killing and host-cell damage. We developed NHAMP, a protein language model (PLM)-fused latent diffusion framework for generating peptides under explicit efficacy-safety constraints. NHAMP maps sequences into ESM-2 embeddings, denoises them with a conditional diffusion model that reuses PLM attention blocks, and decodes latent states with a noise-adapted masked-language-model head. During sampling, classifier-free guidance from antimicrobial, non-hemolytic and dual-property (peptides predicted to be both antimicrobial and non-hemolytic) conditions is accumulated at each denoising step, allowing the model to search more directly for the overlap between activity and safety rather than maximize one property and filter later. Under a shared in silico evaluation pipeline, NHAMP maintained a high predicted AMP-positive rate (83.8%) while increasing the predicted non-hemolytic fraction to 71.4%, with 64.7% of peptides satisfying both predicted criteria in the final benchmark library. In a separate ablation experiment using independently generated sequence sets, triple guidance achieved the highest mean dual-property fraction among the tested guidance settings (60.4 ± 3.05%). Composition, physicochemical and embedding-space analyses showed a consistent shift toward more cationic, less hydrophobic and less aggregation-prone peptides. It should be noted that all performance metrics, including antimicrobial activity (MIC), non-hemolytic probability, and structural confidence, are based strictly on in silico predictions and await further experimental validation. Overall, the results support the idea that introducing safety during denoising, rather than only after generation, can improve the practical yield of AMP libraries for downstream experimental prioritization. Throughout the manuscript, unless experimental validation is explicitly stated, the terms AMP-positive, non-hemolytic and dual-property refer to in silico predictions rather than confirmed biological activity. Scientific contribution This work introduces NHAMP, a protein-language-model–fused latent diffusion framework that applies multi-condition classifier-free guidance by accumulating antimicrobial, predicted non-hemolytic, and dual-property signals at each denoising step, thereby steering sampling toward the narrow overlap between predicted activity and safety. In contrast to common generate-then-filter workflows and diffusion approaches without explicit joint efficacy–safety guidance, NHAMP incorporates this trade-off directly into the generative trajectory and improves the yield of candidates jointly predicted to be antimicrobial and non-hemolytic under a shared in silico evaluation pipeline. Methodologically, the study provides a general and interpretable strategy for integrating competing design objectives during generation rather than only after generation, enabling more efficient prioritization of peptide libraries for subsequent experimental validation.