Machine Learning-Enhanced Engineering of Ultrahigh-Loading Polymer-Drug Assemblies with Programmable ROS-Triggered Release for Tumor Elimination

Overcoming low solubility and bioavailability challenges is critical for enhancing therapeutic efficacy and reducing systemic toxicity of active pharmaceutical ingredients. Here, we describe the rational design of polysulfide nanocarriers that leverage optimized non-covalent polymer-drug interactions—specifically hydrogen bonding, π-interactions, and ion-pairing—to achieve ultrahigh loading capacities (up to ~50 wt.%) across diverse small-molecule therapeutics. Employing machine learning and Bayesian-driven feature refinement, only 4 critical features accurately predicted loading capacity with a mean absolute error of only 2.23% for physically loaded drugs. By systematically tuning polysulfide composition, we created ROS-responsive copolymers capable of stable drug encapsulation and controlled, oxidation-triggered release, tailored for cancer environments. As proof-of-concept, ultrahigh-loaded paclitaxel-polysulfide particles were formulated, showing significantly increased maximum tolerated doses, plasma circulation, tumor retention, and antitumor efficacy compared to clinical standard Taxol. Notably, a Slow-release formulation outperformed a Fast-release formulation, achieving complete tumor eradication in 3-out-of-8 animals in an orthotopic triple negative breast cancer model, emphasizing the importance of curated drug-polymer interactions in nanomedicine. This approach provides critical insights for advancing next-generation delivery platforms and broadening clinical translation possibilities.