Machine learning optimization of polymeric complex coacervate micelles reveals cell-type-specific formulations for enhanced mRNA delivery

The cell-type dependent transfection heterogeneity of polymer-based mRNA delivery remains a fundamental barrier to their broad translation. The field has historically addressed this by optimizing the cationic component itself, refining amine identity, density, and block architecture to maximize transfection. Highly optimized cationic carriers, however, still transfect some cell types well and others poorly, because no single cation chemistry can simultaneously match the internalization and cytoplasmic trafficking biology of every target. Rather than seeking a universally optimal carrier chemistry, we instead sought a tunable formulation framework capable of adapting charge balance and interfacial properties across distinct cellular environments. Here, we report a multicomponent complex coacervate micelle platform in which a cationic oligoamine-functionalized A7 amphiphile, an anionic poly(carboxyethyl acrylate) (CEA) homopolymer, and a neutral-anionic poly(ethylene glycol)-b-poly(carboxyethyl acrylate) (PEG-CEA) diblock are co-assembled into a single carrier, with the anionic and stealth components associating electrostatically with the cationic micelle as external corona species. Closed-loop multi-objective Bayesian optimization (qNEHVI) of the A7/CEA/PEG-CEA formulation simplex over three rounds in HEK293T cells identified a productive ternary region that increased effective efficiency (delivery × viability) from ∼0.35 for the A7 homomicelle to approximately 0.8 for the best mixed micelle. The resulting 35-formulation library was directly screened across six additional cell types (A549, MDA-MB-231, HuH7, OVCAR8, HEK001, and primary patient-derived fibroblasts), where the best mixed micelle outperformed all polymer-based commercial reagents in every cell line. The winning composition was different in every cell line, establishing tunability rather than a single fixed optimum as the platform's central practical advantage. SHapley Additive exPlanations (SHAP) attribution on random forest surrogates then converted these cell-line-specific winners into transferable design rules, identifying which formulation descriptors drive delivery and viability in each cellular context. Together, this work delivers a tunable polymeric transfection platform and a framework for cell-type-aware design of mRNA carriers.