Machine Learning Reveals Amine Type in Polymer Micelles Determines mRNA Binding, In Vitro, and In Vivo Performance for Lung-Selective Delivery

Cationic micelles, composed of amphiphilic block copolymers with polycationic coronas, offer a customizable platform for mRNA delivery. Here, we present a library of 30 cationic micelle nanoparticles (MNPs) formulated from diblock copolymers with reactive poly(pentafluorophenol acrylate) backbones modified with a diverse set of amines. This library systematically varies in nitrogen-based cationic functionalities, exhibiting a spectrum of properties that encompass varied degrees of alkyl substitution (A1-A5), piperazine (A6), oligoamine (A7), guanidinium (A8), hydroxylation (A9-A10) that vary in sidechain volume, substitution pattern, hydrophilicity, and pKa to assess parameter impact on mRNA delivery. In vitro delivery assays using GFP+ mRNA across multiple cell lines reveal that amine sidechain bulk and chemical structure critically affect performance. Using machine learning analysis via SHapley Additive exPlanations (SHAP) on 3,780 experimental data points, we mapped key relationships between amine chemistry and performance metrics, finding that amine-specific binding efficiency was a major determinant of mRNA delivery efficacy, cell viability, and GFP intensity. Micelles with stronger mRNA binding capabilities (A1 and A7) have higher cellular delivery performance, whereas those with intermediate binding tendencies deliver a higher amount of functional mRNA per cell (A2 and A10). This indicates that balancing the binding strength is crucial for performance. Micelles with hydrophobic and bulky pendant groups (A3, A4, and A5) tend to induce necrosis during cellular delivery, highlighting the significance of chemical optimization. A cationic amphiphile identified as A7 displaying a primary and secondary amine, consistently demonstrates the highest GFP expression across various cell types and in vivo achieving high delivery specificity to lung tissue upon intravenous administration. Moreover, we established a strong correlation between in vitro and in vivo performance using Multitask Gaussian Process models, linking amine properties directly to both delivery efficacy and biodistribution. This correlation underscores the predictive power of in vitro models for anticipating in vivo outcomes and highlights chemical amine-dependent optimization as crucial for advancing mRNA delivery vehicle development. Overall, this innovative study integrates advanced data science with experimental design demonstrating the pivotal role of chemical amine identity for targeted mRNA delivery to the lungs.