Generative latent diffusion language modeling yields anti-infective synthetic peptides

Generative artificial intelligence (AI) offers a powerful avenue for peptide design, yet this process remains challenging due to vast sequence space, complex structure-activity relationships, and the need to balance antimicrobial potency with low toxicity. Here, we introduce AMP-Diffusion, a latent diffusion model fine-tuned on antimicrobial peptide (AMP) sequences using embeddings from protein language models (pLMs). AMP-Diffusion enables the rapid discovery of antibiotic candidates by systematically exploring sequence space. We generated 50,000 candidate sequences, filtered and ranked them using our APEX deep learning (DL) model, and synthesized 46 top candidates. These peptides showed broad-spectrum antibacterial activity, including against multidrug-resistant strains, while exhibiting low cytotoxicity. Mechanistic studies revealed membrane permeabilization and depolarization as primary modes of action. In a preclinical mouse model, lead peptides reduced bacterial loads with efficacy comparable to polymyxin B and levofloxacin, with no detectable adverse effects. AMP-Diffusion thus presents a robust platform for designing antibiotics. • AMP-Diffusion designs potent peptides using latent diffusion modeling • 76% of tested peptides kill bacteria with low toxicity • Lead peptides reduce infections in vivo with no adverse effects • Diffusion AI generates diverse, bioactive peptides de novo Antibiotic resistance is accelerating faster than our ability to discover new drugs. Antimicrobial peptides (AMPs) are promising alternatives, but navigating the immense sequence space while meeting potency and safety constraints has been a bottleneck. Here, we introduce AMP-Diffusion, a generative artificial intelligence (AI) platform that uses latent diffusion modeling and protein language embeddings to create biologically relevant results without predefined motifs or structural priors. From 50,000 in silico designs, we synthesized and tested 46 peptides: 76% inhibited bacteria, including multidrug-resistant strains, while showing minimal toxicity. Two lead candidates showed in vivo efficacy in mice that was comparable to clinically used antibiotics. This study illustrates how generative AI can rapidly identify and optimize therapeutic peptides, offering a scalable and generalizable approach to antibiotic development. AMP-Diffusion sets the stage for future platforms that tailor peptides for specific pathogens or therapeutic targets. AMP-Diffusion is a generative artificial intelligence (AI) framework that creates antimicrobial peptides (AMPs) using a latent diffusion model trained on protein language embeddings. Torres et al. show that AI-designed peptides exhibit potent, broad-spectrum antibacterial activity, low toxicity, and in vivo efficacy comparable to standard antibiotics. This work presents a scalable, experimentally validated platform for rational antibiotic discovery, with broader implications for therapeutic peptide design.