From generation to validation: Deep generative models for antimicrobial peptide discovery

The global escalation of antibiotic resistance has renewed interest in antimicrobial peptides (AMPs) as promising alternatives to conventional antibiotics. Although extensive experimental evidence supports their effectiveness against drug-resistant pathogens, identifying new candidates remains time-consuming and labor-intensive. Recent advances in artificial intelligence (AI) are helping to alleviate this bottleneck by accelerating the discovery of AMPs. In particular, generative approaches, including variational autoencoders (VAEs), generative adversarial networks (GANs), and diffusion models, have enabled the de novo design of AMPs with potent activity and reduced toxicity. Despite this progress, key challenges remain, including dataset quality and bias, the calibration and reliability of activity/toxicity prediction, controllable generation under realistic constraints, and rigorous, reproducible validation. This review summarizes recent advances in generative models for AMP discovery, compares different modeling strategies, and outlines mainstream validation methods used to assess peptide efficacy and safety. We provide a systematic comparison and new insights for peptide-centered generative modeling, discuss current limitations and unresolved questions, and propose practical considerations and future directions to guide the development of AMP generation models that are effective, safe, and clinically translatable.