Abstract
The continuous rise of multidrug-resistant pathogens necessitates an urgent need for new antibiotics, yet innovation in antibiotic discovery has largely stalled since the 1980s. In traditional drug development, around 90% of candidate molecules fail at the preclinical stage or phase I of trials due to toxicity or lack of efficacy. Effective antibiotic discovery must overcome a set of microbiological challenges: selective bacterial targeting, penetration of complex cell envelopes, and evasion of diverse resistance mechanisms. Recent advances in deep learning (DL) offer promising opportunities to address these challenges. DL can help identify and characterize new bacterial targets, predict accurate 3-dimensional structures, assess druggability, and discover lead molecules with antibiotic potential. Generative models further enable the de novo design of candidates with optimized pharmacokinetics and safety profiles, potentially resolving long-standing toxicity issues. These technologies streamline labor-intensive screening and boost efficiency in the drug discovery pipeline. However, DL methods need to be applied judiciously. Their effectiveness depends on appropriate model selection, high-quality training data, and careful interpretation of predictions particularly when predicting properties for novel microbial targets. This review provides a timely and critical analysis of DL applications in antibacterial hit discovery through the lens of structural biology, offering structural biologists a road map for integrating these tools into antibiotic discovery workflows to help combat antimicrobial resistance.