Iterative discovery of potent polymeric antibiotics via multi-stage and multi-task learning against antimicrobial resistance.

Drug-resistant bacterial infections pose a serious threat to global health, driving the development of antibacterial strategies beyond classic antibiotics. Host defense peptide mimetic polymeric antibiotics have emerged as promising candidates to combat drug resistance, however, navigating the vast chemical space of polymers remains a significant challenge due to complex structure-activity relationships, while data-driven approaches are further constrained by polymer complexity and scarce labeled data. To address this, we develop PolyCLOVER, a framework that integrates multi-stage self-supervised learning, active learning, and high-throughput experimentation to iteratively discover polymeric antibiotics with potent antibacterial activity and low toxicity. Applied to a combinatorial library of ~100,000 poly(β-amino ester)s, the framework uncovers three lead compounds that self-assemble into stable nanoparticles (SANPs) with minimum inhibitory concentrations of 4 μg/mL and 8 μg/mL against multidrug-resistant S. aureus and A. baumannii, respectively. These SANPs also serve as adjuvant antibiotic carriers, restoring bacterial sensitivity to penicillin G. In vivo studies demonstrate their therapeutic efficacy both as monotherapies and in combination therapies with antibiotics. PolyCLOVER may become a powerful framework for discovery of new polymeric biomaterials without reliance on external datasets.