Antimicrobial resistance, fueled by antibiotic overuse that drives the emergence of resistant strains, is recognized as a global health hazard. It was ranked the third cause of death in 2019, surpassing HIV and malaria [1], creating a pressing need for the discovery of new antimicrobial pharmaceuticals. Antimicrobial peptides (AMPs) are an appealing alternative to known antibiotics [1]. Innate to host defense systems, they combat antibiotic-resistant pathogens, with slower resistance emergence than conventional antibiotics [2]. Despite extensive research, so far, AMPs have faced major hurdles on the way to the clinic [2]. Clinical failures result from low activity, high toxicity, or instability, motivating efforts to design safer, more effective AMPs [3].

Recent years witnessed a tremendous advancement in AI, in particular the development of generative and large language models, revolutionizing the design of drugs [4], proteins [5], as well as AMPs [6, 7, 8, 9]. Since the most recent reviews on AMP discovery cover the principles of traditional machine learning and AI-based methods [9,8,10,11], in this review we focus on classifying the tasks of AI-driven approaches that span the last two years, with particular attention to generative modeling.

We provide a detailed characterization of tasks that the AI methods can perform in AMP discovery, introducing the diverse properties of AMPs and their model representations. We discuss two main categories of AI methods with crucial importance for AMP design: discrimination and generation (Figure 1a). We group the most recent discriminators by their predicted properties and categorize the emerging generators by their modes of unconstrained and analog generation (Figure 1b). We further discuss approaches to the controlled generation of peptides with desired properties (Figure 1c). Moreover, we summarize approaches to the evaluation of AMP discovery, both from the methodological and experimental sides. Finally, we outline unaddressed challenges impeding AMP delivery to the clinic, highlighting the most exciting methodological opportunities for advancement.