Generative Diffusion–Reinforcement Framework with Protein Language Model Conditioning for De Novo Antimicrobial Peptide (AMP) Design and Optimization

Abstract

The rise of antimicrobial resistance (AMR) has created an urgent demand for novel therapeutic agents capable of targeting multi-drug-resistant pathogens. In this work, we propose a Generative AI framework for Antimicrobial Peptide (AMP) design that synergistically combines diffusion modeling, protein language model (PLM) embeddings, and multi-objective reinforcement learning (RL) to generate potent, non-toxic peptide sequences. The framework leverages pretrained sequence embeddings from ESM-2 to capture biochemical and structural priors, while a conditional diffusion prior learns peptide distributions under compositional and biophysical constraints. A reinforcement learning module refines the generative model using a multi-objective reward function balancing antimicrobial potency (Minimum Inhibitory Concentration—MIC), toxicity, stability, and manufacturability. The top-ranked candidates are further filtered via uncertainty-aware active learning, closing the loop with in-vitro validation feedback. Experimental results across benchmark AMP datasets (APD3, CAMP-R4, DBAASP) demonstrate that the proposed model achieves 97.4% sequence validity, 88.9% novelty, and 83.2% Hit@50, outperforming state-of-the-art baselines such as AMPGAN-v2 and Diff-AMP. Visualization of peptide embeddings reveals distinct clusters corresponding to α-helical, β-sheet, and hybrid AMP families, confirming biological interpretability. This integrated generative–reinforcement pipeline establishes a scalable and interpretable foundation for AI-driven antibiotic discovery, accelerating the identification of next-generation AMPs to combat global antimicrobial resistance.

Antimicrobial Peptides (AMPs); Generative Artificial Intelligence; Diffusion Models; Protein Language Models (PLMs); Reinforcement Learning (RL); Multi-Objective Optimization; Minimum Inhibitory Concentration (MIC); Toxicity Prediction; Active Learning; Antimicrobial Resistance (AMR).