AI has designed thousands of potential antibiotics. Will any work? (www.nature.com)

Researchers are increasingly using machine learning and generative AI to design novel antibiotics, producing thousands of candidate molecules in minutes as a faster alternative to traditional, laborious natural-product screening. César de la Fuente’s team trained models on known antibacterial and non-antibacterial compounds and on whole proteomes to identify short peptide fragments with antimicrobial potential. Their generative model produced ~50,000 peptides; a deep-learning ranker prioritized candidates, 46 of which were synthesized. Roughly 35 killed at least one bacterial strain in vitro, most were non-toxic to human embryonic kidney cells, and the top two showed efficacy against Acinetobacter baumannii in mouse models. This work comes amid a worrying rise in resistant infections (US CDC reports a 69% surge in dangerous bacterial infections from 2019–2023; ~1.1 million global deaths annually linked to drug resistance). For the AI/ML community the advance demonstrates generative models’ ability to explore chemical and biological sequence space far beyond what’s practical by hand, dramatically shortening discovery timelines to days or weeks. But significant translational hurdles remain: many AI-designed compounds are chemically unstable, hard or costly to synthesize, and untested in humans, highlighting the need to integrate manufacturability, stability and ADMET constraints into models and to pair in silico design with high-throughput wet-lab validation and medicinal-chemistry optimization before clinical translation.